BIue-Jay commited on
Commit
3102348
·
verified ·
1 Parent(s): e385057

a18cr2: last two checkpoints + eval/deploy files + training records

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/cache_cosmos_reason2.py +60 -0
  2. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/check_reason2_offline.py +26 -0
  3. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/g1_feedback76_h16.py +27 -0
  4. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/groot_chain.py +139 -0
  5. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/h2_body31_hand44_h16.py +24 -0
  6. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/h2_body31_hand44_h32.py +24 -0
  7. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/h2_body31_hand44_planner7_h16.py +25 -0
  8. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/h2w3_body31_hand44_h16.py +24 -0
  9. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/hb_aug_v2.py +163 -0
  10. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/hb_inloop_eval.py +347 -0
  11. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/hb_launch.py +100 -0
  12. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/hb_launch_v.py +102 -0
  13. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/hb_offline_fix.py +24 -0
  14. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/prepare_groot_data.py +191 -0
  15. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/verify_groot_cpu.py +57 -0
  16. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/config.json +90 -0
  17. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/dataset_meta/hb-dataset-receipt.json +47 -0
  18. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/dataset_meta/modality.json +79 -0
  19. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/embodiment_id.json +57 -0
  20. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/experiment_cfg/conf.yaml +251 -0
  21. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/experiment_cfg/config.yaml +249 -0
  22. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/experiment_cfg/dataset_statistics.json +1095 -0
  23. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/experiment_cfg/final_model_config.json +58 -0
  24. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/experiment_cfg/final_processor_config.json +0 -0
  25. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/loss_curve.csv +401 -0
  26. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/model-00001-of-00002.safetensors +3 -0
  27. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/model-00002-of-00002.safetensors +3 -0
  28. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/model.safetensors.index.json +0 -0
  29. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/processor_config.json +1154 -0
  30. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/protocol.json +91 -0
  31. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/statistics.json +0 -0
  32. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/train.log +0 -0
  33. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/trainer_state.json +1234 -0
  34. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/training_args.bin +3 -0
  35. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/training_plan.json +36 -0
  36. a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/wandb_config.json +1 -0
  37. a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/cache_cosmos_reason2.py +60 -0
  38. a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/check_reason2_offline.py +26 -0
  39. a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/g1_feedback76_h16.py +27 -0
  40. a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/groot_chain.py +139 -0
  41. a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/h2_body31_hand44_h16.py +24 -0
  42. a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/h2_body31_hand44_h32.py +24 -0
  43. a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/h2_body31_hand44_planner7_h16.py +25 -0
  44. a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/h2w3_body31_hand44_h16.py +24 -0
  45. a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/hb_aug_v2.py +163 -0
  46. a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/hb_inloop_eval.py +347 -0
  47. a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/hb_launch.py +100 -0
  48. a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/hb_launch_v.py +102 -0
  49. a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/hb_offline_fix.py +24 -0
  50. a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/prepare_groot_data.py +191 -0
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/cache_cosmos_reason2.py ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Put the full gated nvidia/Cosmos-Reason2-2B (pinned 9ce19a1) into the shared HF hub cache $G/hf-home/hub.
2
+
3
+ GR00T N1.7 builds its backbone with Qwen3VLForConditionalGeneration.from_pretrained("nvidia/Cosmos-Reason2-2B") and
4
+ its processor with Qwen3VLProcessor.from_pretrained(same name) before loading the GR00T checkpoint, so every node needs
5
+ the whole repo resolvable by name. groot_node.sh runs with HF_HUB_OFFLINE=1, so nodes read this cache and never need
6
+ the token. refs/main is pinned to the verified revision.
7
+
8
+ Usage (token from stdin, kept in-process only):
9
+ $G/envs/groot-py312/bin/python cache_cosmos_reason2.py $G/hf-home/hub < /root/.hb-secrets/hf_token
10
+ """
11
+ import hashlib
12
+ import json
13
+ import os
14
+ import sys
15
+ import time
16
+
17
+ from huggingface_hub import HfApi, snapshot_download
18
+
19
+ REPO = "nvidia/Cosmos-Reason2-2B"
20
+ REV = "9ce19a195e423419c349abfc86fd07178b230561"
21
+ cache = os.path.abspath(sys.argv[1])
22
+ token = sys.stdin.readline().strip()
23
+ if not token.startswith("hf_"):
24
+ raise SystemExit("token format rejected")
25
+ api = HfApi(endpoint="https://huggingface.co", token=token)
26
+ info = api.model_info(REPO, revision=REV, files_metadata=True)
27
+ assert info.sha == REV, info.sha
28
+ t0 = time.monotonic()
29
+ for attempt in range(1, 21):
30
+ try:
31
+ path = snapshot_download(REPO, revision=REV, cache_dir=cache, token=token, max_workers=8)
32
+ break
33
+ except Exception as error: # transient stalls
34
+ if any(k in str(error).lower() for k in ("gated", "401", "403")):
35
+ raise
36
+ print(f"retry {attempt}: {type(error).__name__}: {str(error)[:200]}", flush=True)
37
+ time.sleep(10)
38
+ else:
39
+ raise SystemExit("download did not finish")
40
+ token = None
41
+ count, size = 0, 0
42
+ for s in info.siblings:
43
+ f = os.path.join(path, s.rfilename)
44
+ assert os.path.isfile(f), s.rfilename
45
+ if s.size is not None:
46
+ assert os.path.getsize(f) == s.size, s.rfilename
47
+ if s.lfs is not None:
48
+ h = hashlib.sha256()
49
+ with open(f, "rb") as stream:
50
+ for chunk in iter(lambda: stream.read(1 << 22), b""):
51
+ h.update(chunk)
52
+ assert h.hexdigest() == s.lfs.sha256, s.rfilename
53
+ count += 1
54
+ size += s.size or 0
55
+ refs = os.path.join(cache, "models--" + REPO.replace("/", "--"), "refs")
56
+ os.makedirs(refs, exist_ok=True)
57
+ with open(os.path.join(refs, "main"), "w") as f:
58
+ f.write(REV)
59
+ print("CACHE_OK", json.dumps({"repo": REPO, "revision": REV, "snapshot": path, "files": count, "bytes": size,
60
+ "seconds": round(time.monotonic() - t0, 1)}), flush=True)
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/check_reason2_offline.py ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """CPU check that GR00T's Cosmos-Reason2-2B pieces load offline from $G/hf-home exactly as groot_node.sh will.
2
+
3
+ Usage: HF_HOME=$G/hf-home HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 $G/envs/groot-py312/bin/python check_reason2_offline.py
4
+ """
5
+ import json
6
+ import os
7
+ import sys
8
+ from pathlib import Path
9
+
10
+ import torch
11
+
12
+ sys.path.insert(0, str(Path(__file__).resolve().parent))
13
+ import hb_offline_fix # noqa: E402,F401
14
+ from transformers import AutoConfig, Qwen3VLForConditionalGeneration, Qwen3VLProcessor
15
+
16
+ assert os.environ.get("HF_HUB_OFFLINE") == "1", "run with HF_HUB_OFFLINE=1"
17
+ name = "nvidia/Cosmos-Reason2-2B"
18
+ torch.set_num_threads(8)
19
+ proc = Qwen3VLProcessor.from_pretrained(name, trust_remote_code=True)
20
+ cfg = AutoConfig.from_pretrained(name, trust_remote_code=True)
21
+ ids = proc.tokenizer("pick up the trash and drop it in the bin")["input_ids"]
22
+ model = Qwen3VLForConditionalGeneration.from_pretrained(name, torch_dtype=torch.bfloat16)
23
+ n = sum(p.numel() for p in model.parameters())
24
+ print("REASON2_OFFLINE_OK", json.dumps({"processor": type(proc).__name__, "tokens": len(ids),
25
+ "model_type": cfg.model_type, "params_M": round(n / 1e6, 1),
26
+ "llm_layers": len(model.model.language_model.layers)}), flush=True)
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/g1_feedback76_h16.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """G1 feedback76 / H16 GR00T modality config (copy of HumanWAM configs/humanwam/groot/feedback76_h16.py @ c76527b).
2
+
3
+ The matching modality.json is generated per dataset from the recorded joint names (prepare_groot_data.py);
4
+ HumanWAM's feedback76_modality.json has right_arm at observation.state[29:36] (left-hand fingers), which is wrong
5
+ for our G1 exports (right arm is [22:29]).
6
+ """
7
+ from gr00t.configs.data.embodiment_configs import register_modality_config
8
+ from gr00t.data.embodiment_tags import EmbodimentTag
9
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
10
+
11
+ feedback76_h16 = {
12
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["ego_view", "chest"]),
13
+ "state": ModalityConfig(
14
+ delta_indices=[0],
15
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "left_hand", "right_hand",
16
+ "projected_gravity"],
17
+ ),
18
+ "action": ModalityConfig(
19
+ delta_indices=list(range(16)),
20
+ modality_keys=["motion_token", "left_hand_joints", "right_hand_joints"],
21
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
22
+ format=ActionFormat.DEFAULT) for _ in range(3)],
23
+ ),
24
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
25
+ }
26
+
27
+ register_modality_config(feedback76_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/groot_chain.py ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Per-node GR00T relay: per-layout batch-size probe (largest that fits), then 20-epoch production per task.
2
+
3
+ Run on every node (nohup), node rank from the DLC env RANK: python groot_chain.py [task ...]
4
+ Default tasks: b6 a3 a7 a18 (basket postponed). A phase succeeds only if all 4 node launchers exit 0.
5
+
6
+ Batch (user rule: as large as memory allows, 2026-11-10): descending CANDIDATES, 4 steps each on all 32 GPUs; a
7
+ candidate fits when all launchers exit 0 AND the peak reserved memory of every rank (written by hb_launch.py to
8
+ logs/peak-<phase>-rank<r>.json) is <= MAX_FRAC of the GPU. Node 0 alone judges each probe and writes a verdict file
9
+ the other nodes wait for, so all nodes take the same decision. One decision per robot layout (g1 / h2).
10
+ Outputs: $G/groot-runs-bmax/ (the earlier per-GPU-64 runs stay in $G/groot-runs/).
11
+ """
12
+ import fcntl
13
+ import json
14
+ import math
15
+ import os
16
+ import shutil
17
+ import subprocess
18
+ import sys
19
+ import time
20
+ from pathlib import Path
21
+
22
+ H = Path(__file__).resolve().parent
23
+ G = H.parent
24
+ RUNS = G / os.environ.get("HB_GROOT_RUNS_NAME", "groot-runs-bmax")
25
+ NODE = int(os.environ["RANK"])
26
+ TASKS = sys.argv[1:] or ["b6", "a3", "a7", "a18"] # basket postponed (user plan 2026-11-10)
27
+ CANDIDATES = [320, 288, 256, 224, 192, 160, 128, 96, 64, 48, 32]
28
+ MAX_FRAC = 0.92
29
+ EPOCHS = 20
30
+ PORT_BASE = int(os.environ.get("HB_GROOT_PORT_BASE", "32500")) # probe PORT_BASE+i, production PORT_BASE+50+k
31
+ RUNS.mkdir(parents=True, exist_ok=True)
32
+ os.environ["HB_GROOT_RUNS"] = str(RUNS) # read by groot_node.sh
33
+ lock = (RUNS / f"chain-node-{NODE}.lock").open("a")
34
+ fcntl.flock(lock, fcntl.LOCK_EX | fcntl.LOCK_NB)
35
+
36
+
37
+ def log(msg):
38
+ line = f"{time.strftime('%Y-%m-%d %H:%M:%S')} node{NODE} {msg}"
39
+ print(line, flush=True)
40
+ with (RUNS / f"chain-node-{NODE}.log").open("a") as f:
41
+ f.write(line + "\n")
42
+
43
+
44
+ def run_phase(task, phase, bs, steps, save, port):
45
+ logs = RUNS / task / "logs"
46
+ exits = [logs / f"{phase}-node-{r}.exit" for r in range(4)]
47
+ if all(p.is_file() for p in exits):
48
+ return all(p.read_text().strip() == "0" for p in exits)
49
+ subprocess.run(["bash", str(H / "groot_node.sh"), task, phase, str(bs), str(steps), str(save), str(port)])
50
+ deadline = time.monotonic() + 600
51
+ while not all(p.is_file() for p in exits):
52
+ if time.monotonic() > deadline:
53
+ log(f"{task}/{phase}: timed out waiting for peer exit files")
54
+ return False
55
+ time.sleep(5)
56
+ ok = all(p.read_text().strip() == "0" for p in exits)
57
+ log(f"{task}/{phase}: bs={bs} steps={steps} -> {'OK' if ok else 'FAIL'}")
58
+ return ok
59
+
60
+
61
+ def peak_frac(task, phase):
62
+ """Max over ranks of peak reserved / total; None when no rank reported."""
63
+ fracs = []
64
+ for f in (RUNS / task / "logs").glob(f"peak-{phase}-rank*.json"):
65
+ try:
66
+ d = json.loads(f.read_text())
67
+ fracs.append(d["peak_reserved_gb"] / d["total_gb"])
68
+ except (ValueError, KeyError, OSError):
69
+ continue
70
+ return (max(fracs), len(fracs)) if fracs else (None, 0)
71
+
72
+
73
+ def probe_verdict(task, cand, port):
74
+ phase = f"probe-bs{cand}"
75
+ verdict = RUNS / task / "logs" / f"{phase}.verdict.json"
76
+ ok = run_phase(task, phase, cand, 4, 1000000, port)
77
+ if NODE == 0 and not verdict.is_file():
78
+ time.sleep(10) # let peak files from the other nodes become visible on NAS
79
+ frac, n = peak_frac(task, phase) if ok else (None, 0)
80
+ fits = bool(ok and frac is not None and frac <= MAX_FRAC)
81
+ verdict.write_text(json.dumps({"per_gpu_batch": cand, "ran_ok": ok, "peak_frac": frac, "ranks_reported": n,
82
+ "max_frac": MAX_FRAC, "fits": fits}) + "\n")
83
+ deadline = time.monotonic() + 600
84
+ while not verdict.is_file():
85
+ if time.monotonic() > deadline:
86
+ log(f"{task}/{phase}: no verdict from node 0")
87
+ return False
88
+ time.sleep(5)
89
+ v = json.loads(verdict.read_text())
90
+ log(f"{task}/{phase}: verdict {v}")
91
+ return v["fits"]
92
+
93
+
94
+ def prune_resume_state(out):
95
+ """Keep weights of every checkpoint but resume state (DeepSpeed optimizer shards, scheduler, RNG) only in the latest."""
96
+ ckpts = sorted((p for p in out.glob("checkpoint-*") if p.name.split("-")[-1].isdigit()),
97
+ key=lambda p: int(p.name.split("-")[-1]))
98
+ for c in ckpts[:-1]:
99
+ for p in list(c.glob("global_step*")) + list(c.glob("optimizer*.pt")) + list(c.glob("scheduler.pt")) \
100
+ + list(c.glob("rng_state*.pth")) + list(c.glob("latest")) + list(c.glob("zero_to_fp32.py")):
101
+ if p.is_dir():
102
+ shutil.rmtree(p, ignore_errors=True)
103
+ else:
104
+ p.unlink(missing_ok=True)
105
+ log(f"{c.name}: resume state dropped (weights kept)")
106
+
107
+
108
+ def batch_for(task):
109
+ layout = "h2" if task in ("a17", "a18") else "g1"
110
+ decision = RUNS / f"batch-decision-{layout}.json"
111
+ if decision.is_file():
112
+ return json.loads(decision.read_text())["per_gpu_batch"]
113
+ for i, cand in enumerate(CANDIDATES):
114
+ if probe_verdict(task, cand, PORT_BASE + i + (20 if layout == "h2" else 0)):
115
+ if NODE == 0:
116
+ decision.write_text(json.dumps({"per_gpu_batch": cand, "global_batch": cand * 32, "probe_task": task,
117
+ "layout": layout, "candidates": CANDIDATES,
118
+ "max_frac": MAX_FRAC}) + "\n")
119
+ return cand
120
+ time.sleep(20)
121
+ log(f"{layout}: no batch size fit; stopping")
122
+ sys.exit(1)
123
+
124
+
125
+ for k, task in enumerate(TASKS):
126
+ bs = batch_for(task)
127
+ log(f"{task}: per-GPU batch {bs} (global {bs * 32})")
128
+ receipt = json.loads((G / "groot-data" / task / "hb-dataset-receipt.json").read_text())
129
+ steps = math.ceil(EPOCHS * receipt["frames"] / (bs * 32))
130
+ save = math.ceil(steps / 2)
131
+ if not run_phase(task, "production", bs, steps, save, PORT_BASE + 50 + k):
132
+ log(f"{task}: production failed; stopping chain")
133
+ sys.exit(1)
134
+ if NODE == 0:
135
+ prune_resume_state(RUNS / task / "production")
136
+ (RUNS / task / "production-complete.json").write_text(json.dumps(
137
+ {"task": task, "frames": receipt["frames"], "epochs": EPOCHS, "per_gpu_batch": bs,
138
+ "global_batch": bs * 32, "max_steps": steps, "save_steps": save}) + "\n")
139
+ log("all tasks done")
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/h2_body31_hand44_h16.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A18 H2 (Sharpa hands) GR00T modality config, same protocol as the A18 Delta-flow run
2
+ (schema a18-h2-sharpa-body31-hand44-state78-chunk16-v1): cameras head+chest, state = 75 joints + projected
3
+ gravity, action = action.body_target (31) + action.hands (44), horizon 16, absolute targets."""
4
+ from gr00t.configs.data.embodiment_configs import register_modality_config
5
+ from gr00t.data.embodiment_tags import EmbodimentTag
6
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
7
+
8
+ h2_body31_hand44_h16 = {
9
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["head", "chest"]),
10
+ "state": ModalityConfig(
11
+ delta_indices=[0],
12
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "head", "left_hand", "right_hand",
13
+ "projected_gravity"],
14
+ ),
15
+ "action": ModalityConfig(
16
+ delta_indices=list(range(16)),
17
+ modality_keys=["body_target", "left_hand_target", "right_hand_target"],
18
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
19
+ format=ActionFormat.DEFAULT) for _ in range(3)],
20
+ ),
21
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
22
+ }
23
+
24
+ register_modality_config(h2_body31_hand44_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/h2_body31_hand44_h32.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """2026-11-15 user: action horizon 32 copy of h2_body31_hand44_h16.py. A18 H2 (Sharpa hands) GR00T modality config, same protocol as the A18 Delta-flow run
2
+ (schema a18-h2-sharpa-body31-hand44-state78-chunk16-v1): cameras head+chest, state = 75 joints + projected
3
+ gravity, action = action.body_target (31) + action.hands (44), horizon 16, absolute targets."""
4
+ from gr00t.configs.data.embodiment_configs import register_modality_config
5
+ from gr00t.data.embodiment_tags import EmbodimentTag
6
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
7
+
8
+ h2_body31_hand44_h32 = {
9
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["head", "chest"]),
10
+ "state": ModalityConfig(
11
+ delta_indices=[0],
12
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "head", "left_hand", "right_hand",
13
+ "projected_gravity"],
14
+ ),
15
+ "action": ModalityConfig(
16
+ delta_indices=list(range(32)),
17
+ modality_keys=["body_target", "left_hand_target", "right_hand_target"],
18
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
19
+ format=ActionFormat.DEFAULT) for _ in range(3)],
20
+ ),
21
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
22
+ }
23
+
24
+ register_modality_config(h2_body31_hand44_h32, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/h2_body31_hand44_planner7_h16.py ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A14 v2 H2 (Sharpa hands) GR00T modality config (H2 training guide hb_training/docs/h2/a14_training_guide.md 4.1):
2
+ cameras head + chest, state78 = 75 joints + projected gravity, action82 = action.body_target (31) + action.hands (44)
3
+ + action.planner (7, the SONIC planner command), horizon 16, absolute targets. Identical to h2_body31_hand44_h16.py
4
+ except the added `planner` action modality."""
5
+ from gr00t.configs.data.embodiment_configs import register_modality_config
6
+ from gr00t.data.embodiment_tags import EmbodimentTag
7
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
8
+
9
+ h2_body31_hand44_planner7_h16 = {
10
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["head", "chest"]),
11
+ "state": ModalityConfig(
12
+ delta_indices=[0],
13
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "head", "left_hand", "right_hand",
14
+ "projected_gravity"],
15
+ ),
16
+ "action": ModalityConfig(
17
+ delta_indices=list(range(16)),
18
+ modality_keys=["body_target", "left_hand_target", "right_hand_target", "planner"],
19
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
20
+ format=ActionFormat.DEFAULT) for _ in range(4)],
21
+ ),
22
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
23
+ }
24
+
25
+ register_modality_config(h2_body31_hand44_planner7_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/h2w3_body31_hand44_h16.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A14 H2 (Sharpa hands) GR00T modality config with THREE cameras head + chest + wrist (user 2026-11-12; the third-person camera is never a model input). Otherwise identical to h2_body31_hand44_h16.py:, same protocol as the A18 Delta-flow run
2
+ (schema a18-h2-sharpa-body31-hand44-state78-chunk16-v1): cameras head+chest, state = 75 joints + projected
3
+ gravity, action = action.body_target (31) + action.hands (44), horizon 16, absolute targets."""
4
+ from gr00t.configs.data.embodiment_configs import register_modality_config
5
+ from gr00t.data.embodiment_tags import EmbodimentTag
6
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
7
+
8
+ h2w3_body31_hand44_h16 = {
9
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["head", "chest", "wrist"]),
10
+ "state": ModalityConfig(
11
+ delta_indices=[0],
12
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "head", "left_hand", "right_hand",
13
+ "projected_gravity"],
14
+ ),
15
+ "action": ModalityConfig(
16
+ delta_indices=list(range(16)),
17
+ modality_keys=["body_target", "left_hand_target", "right_hand_target"],
18
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
19
+ format=ActionFormat.DEFAULT) for _ in range(3)],
20
+ ),
21
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
22
+ }
23
+
24
+ register_modality_config(h2w3_body31_hand44_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/hb_aug_v2.py ADDED
@@ -0,0 +1,163 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Humanoid Benchmark data augmentation v2 (project CLAUDE.md), model-agnostic torch implementation.
2
+
3
+ augment(x, rng, obs_mask=None) takes x: float [T, V, 3, H, W] in [0, 1] (T frames, V cameras) and returns the
4
+ augmented tensor. Each call is one training sample: one of the 15 effects is chosen uniformly and its magnitude
5
+ is scaled by 1.5 around identity; the rest stay at x1. Colour is shared across cameras and time; geometry, light
6
+ patches and white balance are drawn per camera and fixed over time; sensor/optics effects (noise, gaussian blur,
7
+ JPEG, motion blur, sharpen, defocus) apply only to observation frames (obs_mask True).
8
+ Pixel-size parameters are defined at 224 px and scaled by H / 224.
9
+ perturb_joints(q, rng) applies the proprio spec: per-joint bias +-2 deg plus N(0, 0.25 deg) clipped at 3 sigma.
10
+ """
11
+ import io
12
+ import math
13
+
14
+ import numpy as np
15
+ import torch
16
+ import torch.nn.functional as F
17
+ from PIL import Image
18
+ from torchvision.transforms import InterpolationMode
19
+ from torchvision.transforms import functional as TF
20
+
21
+ EFFECTS = ("exposure", "contrast", "saturation", "hue", "gamma", "temperature", "rotation_crop", "noise",
22
+ "gaussian_blur", "jpeg", "motion_blur", "sharpen", "defocus", "light_shadow", "wb_channel")
23
+ CENTER = {"exposure": 0.0, "contrast": 1.0, "saturation": 1.0, "hue": 0.0, "gamma": 1.0, "temperature": 0.0}
24
+ RANGE = {"exposure": (-0.22, 0.22), "contrast": (0.88, 1.12), "saturation": (0.76, 1.24), "hue": (-0.015, 0.015),
25
+ "gamma": (0.89, 1.11), "temperature": (-0.066, 0.066)}
26
+ VERSION = "hb-aug-v2-20261109"
27
+
28
+
29
+ def sample(rng, views, height):
30
+ boost = EFFECTS[int(rng.integers(len(EFFECTS)))]
31
+
32
+ def k(name):
33
+ return 1.5 if boost == name else 1.0
34
+
35
+ px = height / 224.0
36
+ p = {"boost": boost}
37
+ for name, (a, b) in RANGE.items():
38
+ c = CENTER[name]
39
+ p[name] = float(rng.uniform(c + k(name) * (a - c), c + k(name) * (b - c)))
40
+ p["geometry"] = []
41
+ for _ in range(views):
42
+ angle = float(rng.uniform(-1.5 * k("rotation_crop"), 1.5 * k("rotation_crop")))
43
+ p["geometry"].append((angle, 1.0 + abs(angle) * 0.015 + float(rng.uniform(0, 0.02 * k("rotation_crop")))))
44
+ p["noise"] = 4.0 / 255.0 * k("noise") if rng.random() < 0.5 else 0.0
45
+ p["gaussian_blur"] = float(rng.uniform(0.25, 0.65)) * k("gaussian_blur") if rng.random() < 0.15 else 0.0
46
+ if rng.random() < 0.35:
47
+ lo = int(round(100 - k("jpeg") * 30))
48
+ hi = int(round(100 - k("jpeg") * 5))
49
+ p["jpeg"] = int(rng.integers(lo, hi + 1))
50
+ else:
51
+ p["jpeg"] = None
52
+ p["motion_blur"] = ((float(rng.uniform(3, 7)) * k("motion_blur") * px, float(rng.uniform(0, 180)))
53
+ if rng.random() < 0.15 else None)
54
+ p["sharpen"] = float(rng.uniform(0.2, 0.6)) * k("sharpen") if rng.random() < 0.15 else 0.0
55
+ p["defocus"] = float(rng.uniform(1, 2)) * k("defocus") * px if rng.random() < 0.10 else 0.0
56
+ if rng.random() < 0.30:
57
+ g = 0.30 * k("light_shadow")
58
+ p["light_shadow"] = [[(float(rng.uniform(0.1, 0.9)), float(rng.uniform(0.1, 0.9)),
59
+ float(rng.uniform(0.15, 0.35)), float(rng.uniform(1 - g, 1 + g)))
60
+ for _ in range(int(rng.integers(1, 3)))] for _ in range(views)]
61
+ else:
62
+ p["light_shadow"] = None
63
+ if rng.random() < 0.30:
64
+ g = 0.10 * k("wb_channel")
65
+ gains = []
66
+ for _ in range(views):
67
+ row = [1.0, 1.0, 1.0]
68
+ row[int(rng.integers(0, 3))] = float(rng.uniform(1 - g, 1 + g))
69
+ gains.append(row)
70
+ p["wb_channel"] = gains
71
+ else:
72
+ p["wb_channel"] = None
73
+ return p
74
+
75
+
76
+ def _conv(x, kernel):
77
+ n, c, h, w = x.shape
78
+ k = kernel.to(x)[None, None].repeat(c, 1, 1, 1)
79
+ pad = kernel.shape[-1] // 2
80
+ return F.conv2d(F.pad(x, (pad, pad, pad, pad), mode="reflect"), k, groups=c)
81
+
82
+
83
+ def motion_kernel(length, angle_deg):
84
+ size = max(3, int(round(length)) | 1)
85
+ k = torch.zeros(size, size)
86
+ c = size // 2
87
+ th = math.radians(angle_deg)
88
+ for s in np.linspace(-(length - 1) / 2, (length - 1) / 2, 8 * size):
89
+ k[int(round(c - s * math.sin(th))), int(round(c + s * math.cos(th)))] = 1.0
90
+ return k / k.sum()
91
+
92
+
93
+ def disk_kernel(r):
94
+ size = 2 * int(math.ceil(r)) + 1
95
+ c = size // 2
96
+ yy, xx = torch.meshgrid(torch.arange(size), torch.arange(size), indexing="ij")
97
+ k = (((yy - c) ** 2 + (xx - c) ** 2) <= (r + 0.5) ** 2).float()
98
+ return k / k.sum()
99
+
100
+
101
+ def apply(x, p, obs_mask=None, noise_seed=0):
102
+ t, v, c, h, w = x.shape
103
+ x = x.clone().clamp(0, 1)
104
+ for vi, (angle, scale) in enumerate(p["geometry"]):
105
+ x[:, vi] = TF.affine(x[:, vi], angle=angle, translate=[0, 0], scale=scale, shear=[0, 0],
106
+ interpolation=InterpolationMode.BILINEAR)
107
+ if p["light_shadow"] is not None:
108
+ yy, xx = torch.meshgrid(torch.linspace(0, 1, h), torch.linspace(0, 1, w), indexing="ij")
109
+ for vi, blobs in enumerate(p["light_shadow"]):
110
+ m = torch.ones(h, w)
111
+ for cx, cy, s, g in blobs:
112
+ m = m + (g - 1.0) * torch.exp(-((xx - cx) ** 2 + (yy - cy) ** 2) / (2 * s * s))
113
+ x[:, vi] = (x[:, vi] * m.clamp(min=0).to(x)).clamp(0, 1)
114
+ if p["wb_channel"] is not None:
115
+ for vi, g in enumerate(p["wb_channel"]):
116
+ x[:, vi] = (x[:, vi] * torch.tensor(g, dtype=x.dtype).view(1, 3, 1, 1)).clamp(0, 1)
117
+ y = x.reshape(t * v, c, h, w)
118
+ y = (y * float(2.0 ** p["exposure"])).clamp(0, 1)
119
+ d = p["temperature"]
120
+ y = (y * y.new_tensor([1.0 + d, 1.0, 1.0 - d]).view(1, 3, 1, 1)).clamp(0, 1)
121
+ y = ((y - 0.5) * p["contrast"] + 0.5).clamp(0, 1)
122
+ y = TF.adjust_saturation(y, p["saturation"])
123
+ y = TF.adjust_hue(y, p["hue"])
124
+ y = y.clamp(0, 1).pow(p["gamma"])
125
+ x = y.reshape(t, v, c, h, w)
126
+ mask = torch.ones(t, dtype=torch.bool) if obs_mask is None else torch.as_tensor(obs_mask, dtype=torch.bool)
127
+ if bool(mask.any()):
128
+ o = x[mask].reshape(-1, c, h, w)
129
+ if p["gaussian_blur"] > 0:
130
+ o = TF.gaussian_blur(o, [3, 3], [p["gaussian_blur"]] * 2)
131
+ if p["motion_blur"] is not None:
132
+ o = _conv(o, motion_kernel(*p["motion_blur"]))
133
+ if p["defocus"] > 0:
134
+ o = _conv(o, disk_kernel(p["defocus"]))
135
+ if p["sharpen"] > 0:
136
+ o = o + p["sharpen"] * (o - TF.gaussian_blur(o, [5, 5], [1.0, 1.0]))
137
+ o = o.clamp(0, 1)
138
+ if p["noise"] > 0:
139
+ gen = torch.Generator().manual_seed(int(noise_seed))
140
+ o = (o + torch.randn(o.shape, generator=gen, dtype=o.dtype) * p["noise"]).clamp(0, 1)
141
+ if p["jpeg"] is not None:
142
+ for i in range(o.shape[0]):
143
+ buf = io.BytesIO()
144
+ TF.to_pil_image(o[i]).save(buf, format="JPEG", quality=int(p["jpeg"]))
145
+ buf.seek(0)
146
+ with Image.open(buf) as im:
147
+ o[i] = TF.pil_to_tensor(im.convert("RGB")).to(o.dtype).div(255.0)
148
+ x[mask] = o.reshape(-1, v, c, h, w)
149
+ return x.clamp(0, 1)
150
+
151
+
152
+ def augment(x, rng, obs_mask=None):
153
+ p = sample(rng, x.shape[1], x.shape[-2])
154
+ return apply(x, p, obs_mask, noise_seed=int(rng.integers(0, 2 ** 62))), p
155
+
156
+
157
+ def perturb_joints(q, rng):
158
+ q = np.asarray(q, dtype=np.float64)
159
+ d = q.shape[-1]
160
+ bias = rng.uniform(-math.radians(2.0), math.radians(2.0), size=d)
161
+ sigma = math.radians(0.25)
162
+ noise = np.clip(rng.normal(0.0, sigma, size=q.shape), -3 * sigma, 3 * sigma)
163
+ return (q + bias + noise).astype(np.float32)
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/hb_inloop_eval.py ADDED
@@ -0,0 +1,347 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """GR00T in-training open-loop action MSE (2026-11-15 user: every 250 training steps).
2
+
3
+ Imported by hb_groot/hb_launch.py (hb_launch_v.py runs hb_launch.py). It wraps Gr00tTrainer.__init__ so that every
4
+ trainer gets a TrainerCallback. At global_step % HB_INLOOP_EVERY == 0 (default 250; 0 = off) every rank evaluates its
5
+ shard of the benchmark eval cache (eval-action-mse/cache) with the weights being trained right now, exactly as deployed:
6
+ * Gr00tPolicy pipeline: eval processor from <output_dir>/processor (the files every checkpoint ships),
7
+ parse_observation_gr00t inputs, bf16 inputs, model.get_action under inference_mode, processor.decode_action;
8
+ * same protocol and metrics as the post-hoc eval (hb_eval_common): queries every 16 frames, first 16 predicted steps
9
+ vs the recorded actions, nMSE with the training-frame std, per group / per step;
10
+ * split: "heldout" when the task has recorded episodes that are never trained on (B6, trolley, B7), else "train10".
11
+ Training is not changed: module train/eval modes are restored module by module, torch CPU + CUDA (this device), numpy
12
+ and python RNG states are restored after each eval; the flow-matching noise uses a fixed seed (rank, batch) so every
13
+ eval sees the same noise (smoother curve). ZeRO-2 keeps the full bf16 weights on every rank, so ranks run independently;
14
+ partial sums are all-reduced once. Results: <run>/logs/inloop-mse.jsonl (rank 0), one json line per eval, and an
15
+ "HB_INLOOP" log line. Failures never stop training: an exception on any rank disables further evals on all ranks
16
+ (error flag in the same all-reduce, so no rank waits on another).
17
+ Env: HB_INLOOP_EVERY (250), HB_INLOOP_BATCH (8 queries per forward), HB_INLOOP_MAX_QUERIES (0 = all; tests only),
18
+ HB_INLOOP_SPLIT (override), HB_INLOOP_PROCESSOR_DIR (tests only).
19
+ """
20
+ import json
21
+ import os
22
+ import random
23
+ import sys
24
+ import time
25
+ import traceback
26
+ from contextlib import nullcontext
27
+ from pathlib import Path
28
+
29
+ import numpy as np
30
+ import torch
31
+
32
+ EVERY = int(os.environ.get("HB_INLOOP_EVERY", "250"))
33
+ BATCH = int(os.environ.get("HB_INLOOP_BATCH", "8"))
34
+ MAX_Q = int(os.environ.get("HB_INLOOP_MAX_QUERIES", "0"))
35
+ SEED = 1234
36
+ G = Path(__file__).resolve().parent.parent
37
+ EVAL_DIR = G / "eval-action-mse"
38
+ for _p in (str(EVAL_DIR), str(EVAL_DIR / "adapters")):
39
+ if _p not in sys.path:
40
+ sys.path.insert(0, _p)
41
+ import hb_eval_common as E # noqa: E402
42
+ import ad_groot as AG # noqa: E402 (canonical state -> GR00T state groups, same mapping as the post-hoc eval)
43
+
44
+ RANK = int(os.environ.get("RANK", "0"))
45
+ WORLD = int(os.environ.get("WORLD_SIZE", "1"))
46
+
47
+
48
+ def say(*a):
49
+ if RANK == 0:
50
+ print("HB_INLOOP", *a, flush=True)
51
+
52
+
53
+ def _argv(name):
54
+ a = sys.argv
55
+ for i, x in enumerate(a):
56
+ if x == name and i + 1 < len(a):
57
+ return a[i + 1]
58
+ if x.startswith(name + "="):
59
+ return x.split("=", 1)[1]
60
+ return None
61
+
62
+
63
+ def resolve(dataset_path):
64
+ """groot-data/<view> -> (benchmark task, split, view dir) or raises with the reason"""
65
+ view = Path(dataset_path)
66
+ rec = json.loads((view / "hb-dataset-receipt.json").read_text())
67
+ task = E.task_for_source(rec["source"])
68
+ if task is None:
69
+ raise LookupError(f"no benchmark task for source {rec['source']}")
70
+ split = os.environ.get("HB_INLOOP_SPLIT")
71
+ if not split:
72
+ split = "heldout" if (E.split_dir(task, "heldout") / "meta.json").exists() else "train10"
73
+ if not (E.split_dir(task, split) / "meta.json").exists():
74
+ raise FileNotFoundError(f"eval cache {task}/{split} missing (build_cache.py {task} {split})")
75
+ return task, split, view
76
+
77
+
78
+ class InLoopMSE:
79
+ """transformers TrainerCallback (duck-typed: the callback handler only calls the hooks that exist)"""
80
+
81
+ def __init__(self, task, split, view, processor_dir=None):
82
+ self.task, self.split, self.view = task, split, view
83
+ self.processor_dir = processor_dir
84
+ self.disabled = False
85
+ self.ready = False
86
+ self.results = []
87
+
88
+ # ---------------------------------------------------------------- setup (first eval; needs the live model)
89
+ def _setup(self, model, output_dir):
90
+ from transformers import AutoProcessor
91
+ from gr00t.data.embodiment_tags import EmbodimentTag
92
+ from gr00t.policy.gr00t_policy import Gr00tPolicy
93
+ from gr00t.policy.policy import BasePolicy
94
+
95
+ meta = E.load_meta(self.task, self.split)
96
+ self.meta = meta
97
+ self.lay = E.LAYOUTS[meta["layout"]]
98
+ self.std = E.norm_std(E.load_stats(self.task))
99
+ # the Gr00tPolicy of a checkpoint, built around the live model instead of loading weights from disk
100
+ pol = Gr00tPolicy.__new__(Gr00tPolicy)
101
+ BasePolicy.__init__(pol, strict=True)
102
+ pdir = Path(self.processor_dir or os.environ.get("HB_INLOOP_PROCESSOR_DIR") or Path(output_dir) / "processor")
103
+ proc = AutoProcessor.from_pretrained(str(pdir))
104
+ proc.eval()
105
+ proc.validate_strict_action_padding_mask(model.config)
106
+ tag = EmbodimentTag.NEW_EMBODIMENT
107
+ pol.model, pol.processor, pol.embodiment_tag = model, proc, tag
108
+ pol.modality_configs = {k: v for k, v in proc.get_modality_configs()[tag.value].items() if k != "rl_info"}
109
+ pol.collate_fn = proc.collator
110
+ pol.language_key = pol.modality_configs["language"].modality_keys[0]
111
+ self.policy = pol
112
+ self.mod = pol.modality_configs
113
+ self.state_map = AG.H2_STATE if meta["layout"].startswith("h2") else AG.G1_STATE
114
+ self.video_keys = list(self.mod["video"].modality_keys)
115
+ self.state_keys = list(self.mod["state"].modality_keys)
116
+ self.action_keys = list(self.mod["action"].modality_keys)
117
+ missing = [k for k in self.state_keys if k not in self.state_map] + \
118
+ [k for k in self.video_keys if k not in meta["cams"]]
119
+ if missing:
120
+ raise KeyError(f"model expects {missing}, not in the benchmark state / cache cameras {meta['cams']}")
121
+ tj = self.view / "meta/tasks.jsonl"
122
+ self.prompts = {str(r["task_index"]): r["task"] for r in E.jl(tj)} if tj.exists() else {}
123
+ # all queries of the split in a fixed order; this rank keeps every WORLD-th one (images only for those)
124
+ allq = []
125
+ for ep in meta["episodes"]:
126
+ with np.load(E.split_dir(self.task, self.split) / f"ep_{ep:06d}.npz") as z:
127
+ allq += [(ep, qi, int(t)) for qi, t in enumerate(z["query_t"])]
128
+ if MAX_Q:
129
+ allq = allq[:MAX_Q]
130
+ self.n_total = len(allq)
131
+ mine = [q for i, q in enumerate(allq) if i % WORLD == RANK]
132
+ self.eps = list(meta["episodes"])
133
+ self.q = [] # (ep, t, T, state[t], prompt, {cam: img}, gt(16, Da), valid(16))
134
+ by_ep = {}
135
+ for ep, qi, t in mine:
136
+ by_ep.setdefault(ep, []).append((qi, t))
137
+ for ep, lst in by_ep.items():
138
+ with np.load(E.split_dir(self.task, self.split) / f"ep_{ep:06d}.npz") as z:
139
+ state, action = z["state"], z["action"]
140
+ prompt = meta["prompts"][str(int(z["task_index"]))]
141
+ imgs = {c: z[f"img_{c}"] for c in self.video_keys}
142
+ for qi, t in lst:
143
+ gt, valid = E.gt_chunk(action, t)
144
+ self.q.append((ep, t, len(action), state[t].copy(), prompt,
145
+ {c: imgs[c][qi].copy() for c in self.video_keys}, gt, valid))
146
+ self.ready = True
147
+ say(f"ready: task {self.task} split {self.split} queries {self.n_total} (rank {RANK}: {len(self.q)}), "
148
+ f"every {EVERY} steps, batch {BATCH}, cams {self.video_keys}, processor {pdir}")
149
+
150
+ # ---------------------------------------------------------------- one forward on a batch of queries
151
+ def _obs(self, q):
152
+ from gr00t.data.utils import parse_observation_gr00t
153
+ ep, t, T, st, prompt, imgs, _, _ = q
154
+ o = {}
155
+ for k in self.state_keys:
156
+ a, b = self.state_map[k]
157
+ o[f"state.{k}"] = np.asarray(st[a:b], dtype=np.float32)[None]
158
+ for k in self.video_keys:
159
+ o[f"video.{k}"] = imgs[k][None]
160
+ o[self.language_key()] = self.prompts.get("0", prompt) if len(self.prompts) == 1 else prompt
161
+ return parse_observation_gr00t(o, self.mod)
162
+
163
+ def language_key(self):
164
+ return self.policy.language_key
165
+
166
+ @staticmethod
167
+ def _merge(obs_list):
168
+ out = {}
169
+ for m in obs_list[0]:
170
+ out[m] = {}
171
+ for k, v in obs_list[0][m].items():
172
+ vals = [o[m][k] for o in obs_list]
173
+ out[m][k] = np.concatenate(vals, axis=0) if isinstance(v, np.ndarray) else sum(vals, [])
174
+ return out
175
+
176
+ def _predict(self, batch, model):
177
+ obs = self._merge([self._obs(q) for q in batch])
178
+ dtype = next(model.parameters()).dtype
179
+ ctx = torch.autocast("cuda", dtype=torch.bfloat16) if (dtype != torch.bfloat16 and torch.cuda.is_available()) \
180
+ else nullcontext()
181
+ with ctx:
182
+ act, _ = self.policy.get_action(obs)
183
+ return np.concatenate([np.asarray(act[k], dtype=np.float64) for k in self.action_keys], axis=-1)
184
+
185
+ # ---------------------------------------------------------------- the eval
186
+ def evaluate(self, step, model, output_dir):
187
+ t0 = time.time()
188
+ D = self.lay["action_dim"] if self.ready else 0
189
+ err, msg = 0.0, ""
190
+ acc = None
191
+ nq = 0
192
+ infer = 0.0
193
+ modes = [(m, m.training) for m in model.modules()]
194
+ dev = torch.cuda.current_device() if torch.cuda.is_available() else None
195
+ np_state, py_state = np.random.get_state(), random.getstate()
196
+ try:
197
+ with torch.random.fork_rng(devices=[dev] if dev is not None else []):
198
+ if not self.ready:
199
+ self._setup(model, output_dir)
200
+ D = self.lay["action_dim"]
201
+ model.eval()
202
+ acc = E.Accumulator(self.std, self.lay["groups"])
203
+ for bi in range(0, len(self.q), BATCH):
204
+ batch = self.q[bi:bi + BATCH]
205
+ torch.default_generator.manual_seed(SEED + 1000 * RANK + bi)
206
+ if dev is not None:
207
+ torch.cuda.manual_seed(SEED + 1000 * RANK + bi)
208
+ s = time.time()
209
+ pred = self._predict(batch, model)
210
+ infer += time.time() - s
211
+ if pred.shape[0] != len(batch) or pred.shape[1] < E.HORIZON or pred.shape[2] != D:
212
+ raise ValueError(f"prediction {pred.shape}, need ({len(batch)}, >={E.HORIZON}, {D})")
213
+ for q, p in zip(batch, pred):
214
+ acc.add(q[0], p[:E.HORIZON], q[6], q[7])
215
+ nq += 1
216
+ except Exception as e: # noqa: BLE001 - never stop training
217
+ err, msg = 1.0, f"{type(e).__name__}: {e}"
218
+ print(f"HB_INLOOP rank {RANK} step {step} FAILED: {msg}\n{traceback.format_exc()}", file=sys.stderr,
219
+ flush=True)
220
+ finally:
221
+ for m, tr in modes:
222
+ m.training = tr
223
+ np.random.set_state(np_state)
224
+ random.setstate(py_state)
225
+ # ---- one all-reduce of all partial sums (+ error flag); D / eps are identical on every rank once ready
226
+ H = E.HORIZON
227
+ eps = self.eps if self.ready else []
228
+ nE = len(eps)
229
+ vec = np.zeros(3 * D + 1 + 2 * H + 2 * nE + 3)
230
+ if acc is not None and not err:
231
+ o = 0
232
+ vec[o:o + D] = acc.sse_raw; o += D
233
+ vec[o:o + D] = acc.sae_raw; o += D
234
+ vec[o:o + D] = acc.sse_n; o += D
235
+ vec[o] = acc.rows; o += 1
236
+ vec[o:o + H] = acc.step_sse_n; o += H
237
+ vec[o:o + H] = acc.step_rows; o += H
238
+ for j, ep in enumerate(eps):
239
+ s, c = acc.per_episode.get(int(ep), (0.0, 0))
240
+ vec[o + j], vec[o + nE + j] = s, c
241
+ o += 2 * nE
242
+ vec[o], vec[o + 1] = nq, infer
243
+ vec[-1] = err
244
+ tot = self._allreduce(vec)
245
+ if tot is None: # ranks disagree on the vector length (setup failed somewhere) -> stop for good
246
+ self.disabled = True
247
+ self._write(output_dir, {"step": step, "error": "length mismatch across ranks / collective failed"})
248
+ return
249
+ if tot[-1] > 0 or not self.ready:
250
+ self.disabled = True
251
+ self._write(output_dir, {"step": step, "error": msg or "failed on another rank", "ranks_failed": tot[-1]})
252
+ say(f"step {step}: eval failed on {int(tot[-1])} rank(s) -> in-training eval disabled ({msg})")
253
+ return
254
+ o = 0
255
+ acc = E.Accumulator(self.std, self.lay["groups"])
256
+ acc.sse_raw = tot[o:o + D]; o += D
257
+ acc.sae_raw = tot[o:o + D]; o += D
258
+ acc.sse_n = tot[o:o + D]; o += D
259
+ acc.rows = int(round(tot[o])); o += 1
260
+ acc.step_sse_n = tot[o:o + H]; o += H
261
+ acc.step_rows = tot[o:o + H]; o += H
262
+ acc.per_episode = {int(ep): [float(tot[o + j]), int(round(tot[o + nE + j]))]
263
+ for j, ep in enumerate(eps) if tot[o + nE + j] > 0}
264
+ o += 2 * nE
265
+ n_q, infer_sum = int(round(tot[o])), float(tot[o + 1])
266
+ res = acc.result()
267
+ res.pop("nmse_by_dim", None)
268
+ rec = {"step": step, "utc": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()), "task": self.task,
269
+ "split": self.split, "queries": n_q, "ranks": WORLD, "wall_s": round(time.time() - t0, 1),
270
+ "infer_s_per_query": round(infer_sum / max(n_q, 1), 4),
271
+ "denoising_steps": int(getattr(model.action_head, "num_inference_timesteps", -1))
272
+ if hasattr(model, "action_head") else None, "seed": SEED, **res}
273
+ self.results.append(rec)
274
+ self._write(output_dir, rec)
275
+ g = " ".join(f"{k} {v['nmse']:.4f}" for k, v in res["groups"].items())
276
+ say(f"step {step}: {self.task}/{self.split} nMSE {res['nmse']:.4f} ({g}) on {n_q} queries, "
277
+ f"{rec['wall_s']} s")
278
+
279
+ def _allreduce(self, vec):
280
+ import torch.distributed as dist
281
+ if not (dist.is_available() and dist.is_initialized()) or WORLD == 1:
282
+ return vec
283
+ try:
284
+ dev = "cuda" if dist.get_backend() == "nccl" else "cpu"
285
+ n = torch.tensor([float(len(vec))], dtype=torch.float64, device=dev)
286
+ mx = n.clone()
287
+ dist.all_reduce(mx, op=dist.ReduceOp.MAX)
288
+ mn = n.clone()
289
+ dist.all_reduce(mn, op=dist.ReduceOp.MIN)
290
+ if mx.item() != mn.item():
291
+ return None
292
+ t = torch.from_numpy(vec).to(device=dev, dtype=torch.float64)
293
+ dist.all_reduce(t)
294
+ return t.cpu().numpy()
295
+ except Exception as e: # noqa: BLE001
296
+ print(f"HB_INLOOP rank {RANK} all_reduce failed: {e!r}", file=sys.stderr, flush=True)
297
+ return None
298
+
299
+ def _write(self, output_dir, rec):
300
+ if RANK != 0:
301
+ return
302
+ try:
303
+ out = Path(output_dir).parent / "logs" / "inloop-mse.jsonl"
304
+ out.parent.mkdir(parents=True, exist_ok=True)
305
+ with open(out, "a") as f:
306
+ f.write(json.dumps(rec) + "\n")
307
+ except Exception as e: # noqa: BLE001
308
+ print(f"HB_INLOOP write failed: {e!r}", flush=True)
309
+
310
+ # ---------------------------------------------------------------- TrainerCallback hooks
311
+ def on_step_end(self, args, state, control, model=None, **kw):
312
+ step = int(state.global_step)
313
+ if self.disabled or EVERY <= 0 or step <= 0 or step % EVERY or model is None:
314
+ return control
315
+ self.evaluate(step, model, args.output_dir)
316
+ return control
317
+
318
+
319
+ def _install():
320
+ if EVERY <= 0:
321
+ say("disabled (HB_INLOOP_EVERY=0)")
322
+ return
323
+ ds = _argv("--dataset_path") or _argv("--dataset-path")
324
+ if not ds:
325
+ say("disabled: no --dataset_path in argv")
326
+ return
327
+ try:
328
+ task, split, view = resolve(ds)
329
+ except Exception as e: # noqa: BLE001
330
+ say(f"disabled: {e}")
331
+ return
332
+ from transformers import TrainerCallback
333
+ from gr00t.experiment.trainer import Gr00tTrainer
334
+
335
+ cb_cls = type("HBInLoopMSE", (InLoopMSE, TrainerCallback), {})
336
+ _orig_init = Gr00tTrainer.__init__
337
+
338
+ def __init__(self, *a, **k):
339
+ _orig_init(self, *a, **k)
340
+ self.add_callback(cb_cls(task, split, view))
341
+
342
+ Gr00tTrainer.__init__ = __init__
343
+ say(f"installed: task {task} split {split} every {EVERY} steps (view {view.name})")
344
+
345
+
346
+ if os.environ.get("HB_INLOOP_NO_INSTALL") != "1":
347
+ _install()
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/hb_launch.py ADDED
@@ -0,0 +1,100 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """GR00T N1.7 fine-tuning entry with Humanoid Benchmark augmentation v2.
2
+
3
+ Patches Gr00tN1d7Processor in training mode only:
4
+ - images: native random crop / colour jitter / rotation are bypassed (the deterministic eval resize+center-crop is
5
+ used), then hb_aug_v2 is applied jointly to all cameras of the sample;
6
+ - proprio: the body-joint state groups (HB_BODY_STATE_KEYS) get the v2 bias + noise before normalization.
7
+ Then runs the pinned upstream gr00t/experiment/launch_finetune.py unchanged.
8
+ """
9
+ import json
10
+ import os
11
+ import runpy
12
+ import sys
13
+ from pathlib import Path
14
+
15
+ import numpy as np
16
+ import torch
17
+
18
+ HERE = Path(__file__).resolve().parent
19
+ sys.path.insert(0, str(HERE))
20
+ import hb_offline_fix # noqa: E402,F401 (tokenizer load under HF_HUB_OFFLINE=1, see module docstring)
21
+ import hb_aug_v2 as A # noqa: E402
22
+ from gr00t.model.gr00t_n1d7 import processing_gr00t_n1d7 as P # noqa: E402
23
+
24
+ Proc = P.Gr00tN1d7Processor
25
+ BODY_KEYS = [k for k in os.environ.get("HB_BODY_STATE_KEYS", "left_leg,right_leg,waist,left_arm,right_arm").split(",") if k]
26
+ LOG = os.environ.get("HB_AUG_LOG")
27
+ _state = {"rng": None, "count": 0}
28
+
29
+
30
+ def _rng():
31
+ if _state["rng"] is None:
32
+ _state["rng"] = np.random.default_rng([os.getpid(), int.from_bytes(os.urandom(8), "little")])
33
+ return _state["rng"]
34
+
35
+
36
+ _orig_get, _orig_apply, _orig_call = Proc._get_vlm_inputs, Proc._apply_vlm_processing, Proc.__call__
37
+
38
+
39
+ def _get_vlm_inputs(self, image_keys, images, masks, image_transform, language):
40
+ if getattr(self, "training", False):
41
+ self._hb_views = len(image_keys)
42
+ try:
43
+ return _orig_get(self, image_keys, images, masks, self.eval_image_transform, language)
44
+ finally:
45
+ self._hb_views = None
46
+ return _orig_get(self, image_keys, images, masks, image_transform, language)
47
+
48
+
49
+ def _apply_vlm_processing(self, images, language):
50
+ views = getattr(self, "_hb_views", None)
51
+ if views and isinstance(images, torch.Tensor):
52
+ shape = images.shape
53
+ x = images.reshape(-1, views, *shape[1:]).float().div(255.0)
54
+ x, p = A.augment(x, _rng())
55
+ images = (x * 255.0 + 0.5).clamp(0, 255).to(torch.uint8).reshape(shape)
56
+ _state["count"] += 1
57
+ if LOG and _state["count"] in (1, 50):
58
+ with open(f"{LOG}.{os.getpid()}", "a") as f:
59
+ f.write(json.dumps({"count": _state["count"], "views": views, "hw": list(shape[-2:]),
60
+ "boost": p["boost"], "version": A.VERSION}) + "\n")
61
+ return _orig_apply(self, images, language)
62
+
63
+
64
+ def _call(self, messages):
65
+ if getattr(self, "training", False) and BODY_KEYS:
66
+ states = messages[0]["content"].states
67
+ for key in BODY_KEYS:
68
+ if key in states:
69
+ states[key] = A.perturb_joints(states[key], _rng())
70
+ return _orig_call(self, messages)
71
+
72
+
73
+ Proc._get_vlm_inputs = _get_vlm_inputs
74
+ Proc._apply_vlm_processing = _apply_vlm_processing
75
+ Proc.__call__ = _call
76
+
77
+ # 2026-11-15 user: in-training open-loop action MSE every 250 steps (hb_inloop_eval.py adds a TrainerCallback to
78
+ # Gr00tTrainer; HB_INLOOP_EVERY=0 turns it off). It never blocks training: any setup problem only disables it.
79
+ if os.environ.get("HB_PATCH_ONLY") != "1":
80
+ try:
81
+ import hb_inloop_eval # noqa: E402,F401
82
+ except Exception as _e: # noqa: BLE001
83
+ print(f"HB_INLOOP disabled: {_e!r}", flush=True)
84
+
85
+ if os.environ.get("HB_PATCH_ONLY") != "1":
86
+ entry = Path(os.environ["GROOT_SRC"]) / "gr00t/experiment/launch_finetune.py"
87
+ sys.argv = [str(entry)] + sys.argv[1:]
88
+ try:
89
+ runpy.run_path(str(entry), run_name="__main__")
90
+ finally: # batch probe: report this rank's peak reserved memory (groot_chain.py compares it with MAX_FRAC)
91
+ _peak = os.environ.get("HB_PEAK_FILE")
92
+ if _peak and torch.cuda.is_available():
93
+ try:
94
+ _dev = torch.cuda.current_device()
95
+ with open(f"{_peak}-rank{os.environ.get('RANK', '0')}.json", "w") as _f:
96
+ json.dump({"rank": int(os.environ.get("RANK", "0")),
97
+ "peak_reserved_gb": torch.cuda.max_memory_reserved(_dev) / 1e9,
98
+ "total_gb": torch.cuda.get_device_properties(_dev).total_memory / 1e9}, _f)
99
+ except Exception: # noqa: BLE001 - never mask the training result
100
+ pass
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/hb_launch_v.py ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """2026-11-15 user: GR00T variant launcher = hb_launch.py (HB aug v2 + A20 proprio + processor patch, then the pinned
2
+ upstream launch_finetune.py) plus, from environment variables, before the run starts:
3
+ HB_GROOT_MODEL_OVERRIDES JSON {"attn_dropout": 0.0, "diffusion_model_cfg.dropout": 0.0, "num_inference_timesteps": 16,
4
+ ...}: set on config.model (dotted keys go into nested dicts / attributes)
5
+ HB_GROOT_LR_GROUPS JSON {"language": 1e-6, "vision": 1e-5, "projector": 1e-4, "rest": 1e-4}: layer-wise learning
6
+ rates (Psi-Zero style) through Gr00tTrainer.create_optimizer:
7
+ language = backbone ... language_model.* and lm_head
8
+ vision = backbone ... visual.* except the vision->language mergers
9
+ projector = visual.merger.* and visual.deepstack_merger_list.*
10
+ rest = action head (DiT, VL self-attention, VLLN, encoders / decoders)
11
+ weight decay as HF Trainer (no decay on biases / norms); cosine + warmup scale every group.
12
+ Prints HB_LAUNCH_V lines (overrides applied, parameter count and lr per group) on rank 0 for verification."""
13
+ import json
14
+ import os
15
+ import runpy
16
+ import sys
17
+ from pathlib import Path
18
+
19
+ H = Path(__file__).resolve().parent
20
+ sys.path.insert(0, str(H))
21
+ RANK0 = os.environ.get("RANK", "0") == "0"
22
+
23
+
24
+ def say(*a):
25
+ if RANK0:
26
+ print("HB_LAUNCH_V", *a, flush=True)
27
+
28
+
29
+ OV = json.loads(os.environ.get("HB_GROOT_MODEL_OVERRIDES", "{}") or "{}")
30
+ LRG = json.loads(os.environ.get("HB_GROOT_LR_GROUPS", "{}") or "{}")
31
+
32
+ import gr00t.experiment.experiment as E # noqa: E402
33
+
34
+ _run = E.run
35
+
36
+
37
+ def _set(obj, dotted, value):
38
+ parts = dotted.split(".")
39
+ for p in parts[:-1]:
40
+ obj = obj[p] if isinstance(obj, dict) else getattr(obj, p)
41
+ last = parts[-1]
42
+ if isinstance(obj, dict):
43
+ if last not in obj:
44
+ raise KeyError(f"model override {dotted}: no key {last}")
45
+ old = obj[last]
46
+ obj[last] = value
47
+ else:
48
+ if not hasattr(obj, last):
49
+ raise AttributeError(f"model override {dotted}: no attribute {last}")
50
+ old = getattr(obj, last)
51
+ setattr(obj, last, value)
52
+ say(f"override model.{dotted}: {old!r} -> {value!r}")
53
+
54
+
55
+ def run(config):
56
+ for k, v in OV.items():
57
+ _set(config.model, k, v)
58
+ return _run(config)
59
+
60
+
61
+ E.run = run
62
+
63
+ if LRG:
64
+ from transformers import Trainer
65
+ from gr00t.experiment.trainer import Gr00tTrainer
66
+
67
+ def group_of(name):
68
+ if ".language_model." in name or "lm_head" in name:
69
+ return "language"
70
+ if ".visual." in name:
71
+ return "projector" if (".visual.merger." in name or ".visual.deepstack_merger_list." in name) else "vision"
72
+ return "rest"
73
+
74
+ def create_optimizer(self):
75
+ if self.optimizer is not None:
76
+ return self.optimizer
77
+ model = self.model
78
+ decay = set(self.get_decay_parameter_names(model))
79
+ groups, stats = {}, {}
80
+ for n, p in model.named_parameters():
81
+ if not p.requires_grad:
82
+ continue
83
+ g = group_of(n)
84
+ lr = float(LRG.get(g, LRG.get("rest", self.args.learning_rate)))
85
+ wd = self.args.weight_decay if n in decay else 0.0
86
+ groups.setdefault((g, lr, wd), []).append(p)
87
+ s = stats.setdefault(g, [lr, 0, 0])
88
+ s[1] += 1
89
+ s[2] += p.numel()
90
+ opt_cls, opt_kwargs = Trainer.get_optimizer_cls_and_kwargs(self.args, model)
91
+ opt_kwargs.pop("lr", None)
92
+ self.optimizer = opt_cls([{"params": ps, "lr": lr, "weight_decay": wd} for (g, lr, wd), ps in groups.items()],
93
+ lr=self.args.learning_rate, **opt_kwargs)
94
+ for g, (lr, n_t, n_p) in sorted(stats.items()):
95
+ say(f"lr group {g}: lr {lr:g}, {n_t} tensors, {n_p / 1e6:.1f} M params")
96
+ return self.optimizer
97
+
98
+ Gr00tTrainer.create_optimizer = create_optimizer
99
+ say("layer-wise lr groups installed", json.dumps(LRG))
100
+
101
+ say("model overrides", json.dumps(OV))
102
+ runpy.run_path(str(H / "hb_launch.py"), run_name="__main__")
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/hb_offline_fix.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Make tokenizer loading work with HF_HUB_OFFLINE=1 on transformers 4.57.x.
2
+
3
+ PreTrainedTokenizerBase._patch_mistral_regex calls huggingface_hub.model_info() for every hub-named tokenizer with a
4
+ vocab > 100k (to detect Mistral base models). With HF_HUB_OFFLINE=1 that raises OfflineModeIsEnabled even though all
5
+ files are in the cache. GR00T's Cosmos-Reason2-2B / Qwen3-VL tokenizer is not a Mistral tokenizer, so when offline the
6
+ check is skipped and the tokenizer is returned unchanged. Online behaviour is untouched.
7
+ """
8
+ import os
9
+
10
+ from transformers.tokenization_utils_base import PreTrainedTokenizerBase as _PTB
11
+
12
+ if not getattr(_PTB, "_hb_offline_fix", False):
13
+ _orig = _PTB._patch_mistral_regex.__func__
14
+
15
+ def _patch_mistral_regex(cls, tokenizer, *args, **kwargs):
16
+ try:
17
+ return _orig(cls, tokenizer, *args, **kwargs)
18
+ except Exception as error:
19
+ if os.environ.get("HF_HUB_OFFLINE") == "1" or type(error).__name__ == "OfflineModeIsEnabled":
20
+ return tokenizer
21
+ raise
22
+
23
+ _PTB._patch_mistral_regex = classmethod(_patch_mistral_regex)
24
+ _PTB._hb_offline_fix = True
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/prepare_groot_data.py ADDED
@@ -0,0 +1,191 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Build GR00T dataset views (symlinked episodes + rewritten meta + stats) for the G1 tasks.
2
+
3
+ Usage: <groot python> prepare_groot_data.py <task> [<task> ...] (tasks: b6 basket a3 a7)
4
+ B6 always uses the fixed first 100 episodes (0-99) of b6-v5t7-ready-start-20261109.
5
+ """
6
+ import hashlib
7
+ import json
8
+ import os
9
+ import subprocess
10
+ import sys
11
+ from pathlib import Path
12
+
13
+ H = Path(__file__).resolve().parent
14
+ G = H.parent
15
+ P = G.parent
16
+ OUT = G / "groot-data"
17
+ TASKS = {
18
+ "b6": (P / "b6-v5t7-ready-start-60ep-20261109/data/deliveries/g1/b6-v5t7-ready-start-20261109", 100),
19
+ "basket": (P / "basket-current-dynamics-100-v21-b0ff522/data", None), # HF basket-transfer-expert-v1 @current-dynamics-100-v21 b0ff522 (old ed77719 retired)
20
+ "basketv2": (P / "basket-trimv2-c4df4d4/data", None), # HF a20-basket-transfer-pre-reach-trim-expert-v2 @c4df4d40 (user 2026-11-14)
21
+ "trolley": (P / "trolley-push-expert-v1-20261110/data/trolley_push", 100), # HF trolley-push-expert-v1 @ef9024a: fixed first 100 of 103
22
+ "b7": (P / "b7-black-clothes-hamper-v1-1bb002a/data", 100), # HF b7-black-clothes-hamper-expert-v1 @1bb002a, first 100 of 107
23
+ "a6": (P / "a6-door-pull-v1-cf3ee9b/data/door_pull", None), # HF door-pull-expert-v1 @cf3ee9b (A6)
24
+ "b7green": (P / "b7-green-clothes-hamper-v1-ada288c/data", None), # HF b7-green-clothes-hamper-expert-v1 @ada288c, 100 episodes (user 2026-11-13)
25
+ "a7v2": (P / "a7-cabinet-close-v2-aecf070/data", None), # HF a7-cabinet-close-expert-v2 @aecf070, 100 episodes (user 2026-11-14)
26
+ "b8": (P / "b8-table-wipe-v1-3b2e618/data", None), # HF b8-table-wipe-expert-v1 @3b2e618, 100 episodes (user 2026-11-14)
27
+ "b1": (P / "b1-riceball-microwave-v1-36da298/data", None), # HF b1-riceball-microwave-expert-v1 @36da298, all 115 episodes (user 2026-11-14)
28
+ "a10v3": (P / "a10-trolley-push-v3-6e410c5/data", None), # HF a10-trolley-push-expert-v3 @6e410c5, all 148 episodes (user 2026-11-15)
29
+ "a3v3": (P / "a3-fridge-open-v3-24882d5/data", None), # HF a3-fridge-open-expert-v3 @24882d5, all 125 episodes, measured hands (user 2026-11-15)
30
+ "a7v3": (P / "a7-cabinet-close-v3-c2e15b0/data", None), # HF a7-cabinet-close-expert-v3 @c2e15b0, all 100 episodes, measured hands (user 2026-11-15)
31
+ "b6v6": (P / "b6-pedal-bin-v6-ad2f5bf/data", None), # HF b6-pedal-bin @ad2f5bf0 (v6 at the repo root), all 542 episodes (user 2026-11-15)
32
+ "b5": (P / "b5-chair-push-v1-b85b7e9/data", None), # HF b5-chair-push-expert-v1 @b85b7e9, all 100 episodes (user 2026-11-15)
33
+ "a3v4": (P / "a3-fridge-open-v4-6393d4a/data", None), # HF a3-fridge-open-expert-v4 @6393d4a, all 125 episodes (user 2026-11-15: v4 replaces v3)
34
+ "a7v4": (P / "a7-cabinet-close-v4-90cc2b5/data", None), # HF a7-cabinet-close-expert-v4 @90cc2b5, all 100 episodes (user 2026-11-15: v4 replaces v3)
35
+ "b7greenv2": (P / "b7-green-clothes-hamper-v2-c12757e/data", None), # HF b7-green-clothes-hamper @c12757e (v2 at the repo root), all 120 episodes (user 2026-11-14)
36
+ "a10v2": (P / "a10-trolley-push-v2-d3abb4c/data", None), # HF a10-trolley-push-expert-v2 @d3abb4c, all 150 episodes (user 2026-11-13)
37
+ "a1": (P / "a1-ground-object-pickup-v1-f47f07e/data", None), # HF a1-ground-object-pickup-expert-v1 @f47f07e (A1)
38
+ "c6t": (P / "c6-test8-06c231d/data", None), # c6 8-episode adaptation test only
39
+ "c6": (P / "c6-bathroom-door-fetch-v1-3afba69/data", None), # HF c6-bathroom-door-fetch-expert-v1 @3afba69 (C6)
40
+ "a3": (P / "a3-a7-20261109/a3-fridge-open-expert-v1/780fcf1e4bd7db947a8c2d8f7b92e61f24f71b45", None),
41
+ "a7": (P / "a3-a7-20261109/a7-cabinet-close-expert-v1/1935a98e8f55d4d1d2d1615ddb97c2679713595c", None),
42
+ "a18": (P / "a18-h2-dataset-audit-8e956a10/dataset", None),
43
+ "a17": (P / "a17-unscrew-bottle-v1-407e6e2/data", None), # HF a17-unscrew-bottle-expert-v1 @407e6e2
44
+ "a18c2": (P / "a18-h2-dataset-audit-8e956a10/dataset", None), # A18, 2 cameras head + chest (user 2026-11-14 18:2x)
45
+ "a17c2": (P / "a17-unscrew-bottle-v1-407e6e2/data", None), # A17, 2 cameras head + chest (user 2026-11-14 18:2x)
46
+ "a18cr2": (P / "a18-pour-water-cropped-v2-44cdd57/data", None), # HF a18-pour-water-expert-cropped-v2 @44cdd57, head + chest (user 2026-11-15)
47
+ "a17cr2": (P / "a17-unscrew-bottle-cropped-v2-3520b70/data", None), # HF a17-unscrew-bottle-expert-cropped-v2 @3520b70, head + chest (user 2026-11-15)
48
+ "a14": (P / "a14-unscrew-bulb-v1-b4390fc/data", None), # HF a14-unscrew-bulb-expert-v1 @b4390fc (H2, 3 cameras)
49
+ "a14v2": (P / "a14-unscrew-bulb-v2-8411a6d/data", None), # HF a14-unscrew-bulb-expert-v2 @8411a6da: + action.planner (H2 guide 2026-11-14)
50
+ }
51
+ CONFIG = {"a18": "h2w3_body31_hand44_h16.py", "a17": "h2w3_body31_hand44_h16.py", "a14": "h2w3_body31_hand44_h16.py",
52
+ "a18c2": "h2_body31_hand44_h16.py", "a17c2": "h2_body31_hand44_h16.py",
53
+ "a18cr2": "h2_body31_hand44_h16.py", "a17cr2": "h2_body31_hand44_h16.py",
54
+ "a14v2": "h2_body31_hand44_planner7_h16.py"} # 2026-11-14 H2 guide: action82 incl. the planner command
55
+ # model-input cameras per H2 task (user 2026-11-12 22:0x: EVERY H2 task uses head + chest + wrist; third_person never)
56
+ CAMS = {"a18": ("head", "chest", "wrist"), "a17": ("head", "chest", "wrist"), "a14": ("head", "chest", "wrist"),
57
+ "a18c2": ("head", "chest"), "a17c2": ("head", "chest"), "a18cr2": ("head", "chest"), "a17cr2": ("head", "chest"), "a14v2": ("head", "chest")} # 2026-11-14: 2-camera re-train (user)
58
+ PLANNER = {"a14v2"} # tasks whose action adds action.planner (7) after body_target + hands (H2 guide a14 2.2)
59
+ H2_GROUPS = {"left_leg": (0, 6), "right_leg": (6, 12), "waist": (12, 15), "left_arm": (15, 22), "right_arm": (22, 29),
60
+ "head": (29, 31), "left_hand": (31, 53), "right_hand": (53, 75)}
61
+
62
+
63
+ def h2_modality(info, cams=("head", "chest")):
64
+ f = info["features"]
65
+ names = f["observation.state"]["names"]
66
+ assert len(names) == 75 and names[:31] == f["action.body_target"]["names"] and names[31:] == f["action.hands"]["names"]
67
+ assert f["observation.projected_gravity"]["shape"] == [3]
68
+ state = {g: {"original_key": "observation.state", "start": a, "end": b} for g, (a, b) in H2_GROUPS.items()}
69
+ state["projected_gravity"] = {"original_key": "observation.projected_gravity", "start": 0, "end": 3}
70
+ return {"state": state,
71
+ "action": {"body_target": {"original_key": "action.body_target", "start": 0, "end": 31},
72
+ "left_hand_target": {"original_key": "action.hands", "start": 0, "end": 22},
73
+ "right_hand_target": {"original_key": "action.hands", "start": 22, "end": 44}},
74
+ "video": {c: {"original_key": f"observation.images.{c}"} for c in cams},
75
+ "annotation": {"human.task_description": {"original_key": "task_index"}}}
76
+ BODY = ["left_hip_pitch_joint", "left_hip_roll_joint", "left_hip_yaw_joint", "left_knee_joint", "left_ankle_pitch_joint",
77
+ "left_ankle_roll_joint", "right_hip_pitch_joint", "right_hip_roll_joint", "right_hip_yaw_joint",
78
+ "right_knee_joint", "right_ankle_pitch_joint", "right_ankle_roll_joint", "waist_yaw_joint", "waist_roll_joint",
79
+ "waist_pitch_joint", "left_shoulder_pitch_joint", "left_shoulder_roll_joint", "left_shoulder_yaw_joint",
80
+ "left_elbow_joint", "left_wrist_roll_joint", "left_wrist_pitch_joint", "left_wrist_yaw_joint",
81
+ "right_shoulder_pitch_joint", "right_shoulder_roll_joint", "right_shoulder_yaw_joint", "right_elbow_joint",
82
+ "right_wrist_roll_joint", "right_wrist_pitch_joint", "right_wrist_yaw_joint"]
83
+ GROUPS = {"left_leg": (0, 6), "right_leg": (6, 12), "waist": (12, 15), "left_arm": (15, 22), "right_arm": (22, 29)}
84
+
85
+
86
+ def jl(path):
87
+ return [json.loads(x) for x in Path(path).read_text().splitlines() if x.strip()]
88
+
89
+
90
+ def wl(path, rows):
91
+ Path(path).write_text("".join(json.dumps(r) + "\n" for r in rows))
92
+
93
+
94
+ def build(task):
95
+ src, limit = TASKS[task]
96
+ info = json.loads((src / "meta/info.json").read_text())
97
+ names = info["features"]["observation.state"]["names"]
98
+ if task in CONFIG:
99
+ idx = list(range(31))
100
+ else:
101
+ idx = [names.index(n) for n in BODY]
102
+ for g, (a, b) in GROUPS.items():
103
+ assert idx[a:b] == list(range(idx[a], idx[a] + (b - a))), (task, g, idx[a:b])
104
+ for key, dim in (("action.motion_token", 64), ("observation.inspire_left_hand_state", 6),
105
+ ("observation.inspire_right_hand_state", 6), ("observation.projected_gravity", 3),
106
+ ("observation.images.ego_view", None), ("observation.images.chest", None)):
107
+ assert key in info["features"], (task, key)
108
+ if dim:
109
+ assert info["features"][key]["shape"] == [dim], (task, key)
110
+ episodes = jl(src / "meta/episodes.jsonl")
111
+ stats = jl(src / "meta/episodes_stats.jsonl")
112
+ n = len(episodes) if limit is None else limit
113
+ episodes, stats = episodes[:n], stats[:n]
114
+ assert [e["episode_index"] for e in episodes] == list(range(n)) == [s["episode_index"] for s in stats]
115
+ frames = sum(int(e["length"]) for e in episodes)
116
+ dst = OUT / task
117
+ if dst.exists():
118
+ raise SystemExit(f"{dst} exists")
119
+ (dst / "meta").mkdir(parents=True)
120
+ vkeys = [k for k, f in info["features"].items() if f.get("dtype") == "video"]
121
+ for e in episodes:
122
+ i = e["episode_index"]
123
+ chunk = i // int(info["chunks_size"])
124
+ rel = info["data_path"].format(episode_chunk=chunk, episode_index=i)
125
+ (dst / rel).parent.mkdir(parents=True, exist_ok=True)
126
+ os.symlink(src / rel, dst / rel)
127
+ for k in vkeys:
128
+ relv = info["video_path"].format(episode_chunk=chunk, video_key=k, episode_index=i)
129
+ (dst / relv).parent.mkdir(parents=True, exist_ok=True)
130
+ os.symlink(src / relv, dst / relv)
131
+ if task in CONFIG:
132
+ # stats.py computes stats for every low-dim feature in info.json and cannot handle e.g. observation.tactile
133
+ # [10,24,24]; keep only what the H2 modality uses (parquet columns stay untouched behind the symlinks).
134
+ keep = {"observation.state", "observation.projected_gravity", "action.body_target", "action.hands",
135
+ "timestamp", "frame_index", "episode_index", "index", "task_index"}
136
+ keep |= {f"observation.images.{c}" for c in CAMS[task]}
137
+ keep |= {"action.planner"} if task in PLANNER else set()
138
+ info["features"] = {k: v for k, v in info["features"].items() if k in keep}
139
+ vkeys = [k for k in vkeys if k in keep]
140
+ stats = [{**s, "stats": {k: v for k, v in s["stats"].items() if k in keep}} if "stats" in s else s
141
+ for s in stats]
142
+ info.update(total_episodes=n, total_frames=frames, total_videos=n * len(vkeys),
143
+ total_chunks=(n - 1) // int(info["chunks_size"]) + 1, splits={"train": f"0:{n}"})
144
+ (dst / "meta/info.json").write_text(json.dumps(info, indent=2))
145
+ wl(dst / "meta/episodes.jsonl", episodes)
146
+ wl(dst / "meta/episodes_stats.jsonl", stats)
147
+ (dst / "meta/tasks.jsonl").write_text((src / "meta/tasks.jsonl").read_text())
148
+ if task in CONFIG:
149
+ modality = h2_modality(info, CAMS[task])
150
+ if task in PLANNER:
151
+ assert info["features"]["action.planner"]["shape"] == [7]
152
+ modality["action"]["planner"] = {"original_key": "action.planner", "start": 0, "end": 7}
153
+ else:
154
+ state = {g: {"original_key": "observation.state", "start": idx[a], "end": idx[a] + (b - a)}
155
+ for g, (a, b) in GROUPS.items()}
156
+ state.update(left_hand={"original_key": "observation.inspire_left_hand_state", "start": 0, "end": 6},
157
+ right_hand={"original_key": "observation.inspire_right_hand_state", "start": 0, "end": 6},
158
+ projected_gravity={"original_key": "observation.projected_gravity", "start": 0, "end": 3})
159
+ modality = {"state": state,
160
+ "action": {"motion_token": {"original_key": "action.motion_token", "start": 0, "end": 64},
161
+ "left_hand_joints": {"original_key": "observation.inspire_left_hand_state", "start": 0, "end": 6},
162
+ "right_hand_joints": {"original_key": "observation.inspire_right_hand_state", "start": 0, "end": 6}},
163
+ "video": {"ego_view": {"original_key": "observation.images.ego_view"},
164
+ "chest": {"original_key": "observation.images.chest"}},
165
+ "annotation": {"human.task_description": {"original_key": "task_index"}}}
166
+ (dst / "meta/modality.json").write_text(json.dumps(modality, indent=2))
167
+ py = sys.executable
168
+ src_g = G / "HumanWAM/third_party/Isaac-GR00T"
169
+ env = dict(os.environ, PYTHONPATH=f"{src_g}:{H}")
170
+ done = None
171
+ for tag in ("NEW_EMBODIMENT", "new_embodiment"):
172
+ r = subprocess.run([py, str(src_g / "gr00t/data/stats.py"), "--dataset-path", str(dst), "--embodiment-tag", tag,
173
+ "--modality-config-path", str(H / CONFIG.get(task, "g1_feedback76_h16.py"))], env=env, cwd=src_g,
174
+ capture_output=True, text=True)
175
+ if r.returncode == 0:
176
+ done = tag
177
+ break
178
+ last = r.stdout[-1500:] + r.stderr[-1500:]
179
+ if done is None:
180
+ raise SystemExit(f"stats failed for {task}:\n{last}")
181
+ assert (dst / "meta/stats.json").is_file()
182
+ receipt = dict(task=task, source=str(src), episodes=n, frames=frames, video_keys=vkeys,
183
+ state_dim_raw=len(names), body_index=idx, stats_tag=done,
184
+ modality_sha256=hashlib.sha256((dst / "meta/modality.json").read_bytes()).hexdigest(),
185
+ stats_sha256=hashlib.sha256((dst / "meta/stats.json").read_bytes()).hexdigest())
186
+ (dst / "hb-dataset-receipt.json").write_text(json.dumps(receipt, indent=2) + "\n")
187
+ print("DATASET_OK", json.dumps(receipt)[:400], flush=True)
188
+
189
+
190
+ for t in sys.argv[1:]:
191
+ build(t)
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/code/hb_groot/verify_groot_cpu.py ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """CPU check of the GR00T data path for one task view: loader shapes, v2 augmentation, processor patch import."""
2
+ import json
3
+ import os
4
+ import sys
5
+ from pathlib import Path
6
+
7
+ import numpy as np
8
+ import torch
9
+
10
+ H = Path(__file__).resolve().parent
11
+ sys.path.insert(0, str(H))
12
+ os.environ["HB_PATCH_ONLY"] = "1"
13
+ import hb_launch # noqa: E402,F401 (applies the processor patch)
14
+ import hb_aug_v2 as A # noqa: E402
15
+ from gr00t.data.dataset.lerobot_episode_loader import LeRobotEpisodeLoader # noqa: E402
16
+ from gr00t.model.gr00t_n1d7.processing_gr00t_n1d7 import Gr00tN1d7Processor # noqa: E402
17
+
18
+ view = Path(sys.argv[1])
19
+ if view.name == "a14v2": # 2026-11-14 H2 guide: head + chest, action82 = body_target + hands + planner
20
+ import h2_body31_hand44_planner7_h16 as C # noqa: E402
21
+ cfg = C.h2_body31_hand44_planner7_h16
22
+ elif view.name in ("a18c2", "a17c2", "a18cr2", "a17cr2"): # 2026-11-14: H2 2-camera re-train (head + chest)
23
+ import h2_body31_hand44_h16 as C # noqa: E402
24
+ cfg = next(v for k, v in vars(C).items() if isinstance(v, dict) and "video" in v)
25
+ elif view.name in ("a14", "a18", "a17"): # every H2 task: 3 cameras head + chest + wrist (user 2026-11-12)
26
+ import h2w3_body31_hand44_h16 as C # noqa: E402
27
+ cfg = C.h2w3_body31_hand44_h16
28
+ else:
29
+ import g1_feedback76_h16 as C # noqa: E402
30
+ cfg = C.feedback76_h16
31
+ loader = LeRobotEpisodeLoader(view, cfg)
32
+ df = loader[0]
33
+ cols = list(df.columns)
34
+ row = df.iloc[20]
35
+ out = {"episodes": len(loader), "columns": cols[:40], "rows_ep0": len(df)}
36
+ state_dims = {c: int(np.asarray(row[c]).size) for c in cols if c.startswith("state.")}
37
+ action_dims = {c: int(np.asarray(row[c]).size) for c in cols if c.startswith("action.")}
38
+ out.update(state_dims=state_dims, action_dims=action_dims, state_total=sum(state_dims.values()),
39
+ action_total=sum(action_dims.values()))
40
+ vids = [c for c in cols if c.startswith("video.")]
41
+ imgs = [np.asarray(row[c]) for c in vids]
42
+ out["video"] = {c: list(i.shape) for c, i in zip(vids, imgs)}
43
+ x = torch.stack([torch.from_numpy(i).permute(2, 0, 1).float().div(255) for i in imgs])[None]
44
+ rng = np.random.default_rng(0)
45
+ boosts = {}
46
+ for _ in range(60):
47
+ y, p = A.augment(x, rng)
48
+ boosts[p["boost"]] = boosts.get(p["boost"], 0) + 1
49
+ assert y.shape == x.shape and torch.isfinite(y).all() and float(y.min()) >= 0 and float(y.max()) <= 1
50
+ out["aug_boost_counts_60"] = boosts
51
+ out["mean_abs_change_last"] = round(float((y - x).abs().mean()), 4)
52
+ q = np.zeros((1, 29))
53
+ dq = A.perturb_joints(q, rng)
54
+ out["proprio_max_abs_deg"] = round(float(np.degrees(np.abs(dq).max())), 3)
55
+ out["patched"] = Gr00tN1d7Processor._get_vlm_inputs.__name__ == "_get_vlm_inputs" and \
56
+ Gr00tN1d7Processor.__call__.__module__ == "hb_launch"
57
+ print("CPU_VERIFY", json.dumps(out), flush=True)
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/config.json ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "action_horizon": 40,
3
+ "add_pos_embed": true,
4
+ "apply_sincos_state_encoding": false,
5
+ "architectures": [
6
+ "Gr00tN1d7"
7
+ ],
8
+ "attn_dropout": 0.2,
9
+ "attn_implementation": null,
10
+ "backbone_embedding_dim": 2048,
11
+ "backbone_trainable_params_fp32": true,
12
+ "color_jitter_params": {
13
+ "brightness": 0.3,
14
+ "contrast": 0.4,
15
+ "hue": 0.08,
16
+ "saturation": 0.5
17
+ },
18
+ "crop_fraction": 0.95,
19
+ "diffusion_model_cfg": {
20
+ "attention_head_dim": 48,
21
+ "dropout": 0.2,
22
+ "final_dropout": true,
23
+ "interleave_self_attention": true,
24
+ "norm_type": "ada_norm",
25
+ "num_attention_heads": 32,
26
+ "num_layers": 32,
27
+ "output_dim": 1024,
28
+ "positional_embeddings": null
29
+ },
30
+ "dtype": "bfloat16",
31
+ "exclude_state": false,
32
+ "formalize_language": true,
33
+ "hidden_size": 1024,
34
+ "image_crop_size": [
35
+ 230,
36
+ 230
37
+ ],
38
+ "image_target_size": [
39
+ 256,
40
+ 256
41
+ ],
42
+ "letter_box_transform": false,
43
+ "load_bf16": false,
44
+ "max_action_dim": 132,
45
+ "max_num_embodiments": 32,
46
+ "max_seq_len": 1024,
47
+ "max_state_dim": 132,
48
+ "model_dtype": "bfloat16",
49
+ "model_name": "nvidia/Cosmos-Reason2-2B",
50
+ "model_type": "Gr00tN1d7",
51
+ "noise_beta_alpha": 1.5,
52
+ "noise_beta_beta": 1.0,
53
+ "noise_s": 0.999,
54
+ "num_inference_timesteps": 4,
55
+ "num_timestep_buckets": 1000,
56
+ "random_history_crop": true,
57
+ "random_rotation_angle": 0,
58
+ "reproject_vision": false,
59
+ "rtc_ramp_rate": 6.0,
60
+ "select_layer": 16,
61
+ "shortest_image_edge": 256,
62
+ "state_dropout_prob": 0.0,
63
+ "state_gaussian_noise_std": 0.0,
64
+ "strict_action_padding_mask": false,
65
+ "strict_vl_padding_mask": false,
66
+ "transformers_version": "4.57.3",
67
+ "tune_diffusion_model": true,
68
+ "tune_linear": true,
69
+ "tune_llm": true,
70
+ "tune_projector": true,
71
+ "tune_top_llm_layers": 0,
72
+ "tune_visual": true,
73
+ "tune_vlln": true,
74
+ "use_albumentations": true,
75
+ "use_alternate_vl_dit": true,
76
+ "use_flash_attention": true,
77
+ "use_future_tokens": false,
78
+ "use_mean_std": false,
79
+ "use_percentiles": true,
80
+ "use_vl_self_attention": true,
81
+ "use_vlln": true,
82
+ "vl_self_attention_cfg": {
83
+ "attention_head_dim": 64,
84
+ "dropout": 0.2,
85
+ "final_dropout": true,
86
+ "num_attention_heads": 32,
87
+ "num_layers": 4,
88
+ "positional_embeddings": null
89
+ }
90
+ }
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/dataset_meta/hb-dataset-receipt.json ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "a18cr2",
3
+ "source": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/a18-pour-water-cropped-v2-44cdd57/data",
4
+ "episodes": 100,
5
+ "frames": 66850,
6
+ "video_keys": [
7
+ "observation.images.head",
8
+ "observation.images.chest"
9
+ ],
10
+ "state_dim_raw": 75,
11
+ "body_index": [
12
+ 0,
13
+ 1,
14
+ 2,
15
+ 3,
16
+ 4,
17
+ 5,
18
+ 6,
19
+ 7,
20
+ 8,
21
+ 9,
22
+ 10,
23
+ 11,
24
+ 12,
25
+ 13,
26
+ 14,
27
+ 15,
28
+ 16,
29
+ 17,
30
+ 18,
31
+ 19,
32
+ 20,
33
+ 21,
34
+ 22,
35
+ 23,
36
+ 24,
37
+ 25,
38
+ 26,
39
+ 27,
40
+ 28,
41
+ 29,
42
+ 30
43
+ ],
44
+ "stats_tag": "NEW_EMBODIMENT",
45
+ "modality_sha256": "50d8651783c576fa6f531b1a1c11add620ba1cd3f5752543afd8bcf0b0590bb3",
46
+ "stats_sha256": "9baa6f2ae4d2b3cced4015eff350c6a2b493d899a76d3d21c94cdd7a5db529bd"
47
+ }
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/dataset_meta/modality.json ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "state": {
3
+ "left_leg": {
4
+ "original_key": "observation.state",
5
+ "start": 0,
6
+ "end": 6
7
+ },
8
+ "right_leg": {
9
+ "original_key": "observation.state",
10
+ "start": 6,
11
+ "end": 12
12
+ },
13
+ "waist": {
14
+ "original_key": "observation.state",
15
+ "start": 12,
16
+ "end": 15
17
+ },
18
+ "left_arm": {
19
+ "original_key": "observation.state",
20
+ "start": 15,
21
+ "end": 22
22
+ },
23
+ "right_arm": {
24
+ "original_key": "observation.state",
25
+ "start": 22,
26
+ "end": 29
27
+ },
28
+ "head": {
29
+ "original_key": "observation.state",
30
+ "start": 29,
31
+ "end": 31
32
+ },
33
+ "left_hand": {
34
+ "original_key": "observation.state",
35
+ "start": 31,
36
+ "end": 53
37
+ },
38
+ "right_hand": {
39
+ "original_key": "observation.state",
40
+ "start": 53,
41
+ "end": 75
42
+ },
43
+ "projected_gravity": {
44
+ "original_key": "observation.projected_gravity",
45
+ "start": 0,
46
+ "end": 3
47
+ }
48
+ },
49
+ "action": {
50
+ "body_target": {
51
+ "original_key": "action.body_target",
52
+ "start": 0,
53
+ "end": 31
54
+ },
55
+ "left_hand_target": {
56
+ "original_key": "action.hands",
57
+ "start": 0,
58
+ "end": 22
59
+ },
60
+ "right_hand_target": {
61
+ "original_key": "action.hands",
62
+ "start": 22,
63
+ "end": 44
64
+ }
65
+ },
66
+ "video": {
67
+ "head": {
68
+ "original_key": "observation.images.head"
69
+ },
70
+ "chest": {
71
+ "original_key": "observation.images.chest"
72
+ }
73
+ },
74
+ "annotation": {
75
+ "human.task_description": {
76
+ "original_key": "task_index"
77
+ }
78
+ }
79
+ }
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/embodiment_id.json ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "robocasa_panda_omron": 13,
3
+ "oxe_droid": 17,
4
+ "oxe_fractal": 18,
5
+ "oxe_language_table": 19,
6
+ "oxe_bridge": 20,
7
+ "unknown": 22,
8
+ "gr1_unified": 20,
9
+ "agibot": 26,
10
+ "sim_behavior_r1_pro": 23,
11
+ "xdof": 24,
12
+ "xdof_oss_data": 25,
13
+ "unitree_g1_full_body_with_waist_height_nav_cmd": 25,
14
+ "real_r1_pro_sharpa": 27,
15
+ "real_r1_pro_sharpa_add_view": 27,
16
+ "real_r1_pro_sharpa_relative_arm_joint": 26,
17
+ "real_r1_pro_sharpa_delta_eef": 26,
18
+ "real_r1_pro_sharpa_absolute_eef": 26,
19
+ "real_r1_pro_sharpa_meanstd": 26,
20
+ "real_r1_pro_sharpa_relative_eef": 26,
21
+ "real_r1_pro_sharpa_relative_eef_add_view": 26,
22
+ "real_r1_pro_sharpa_relative_eef_relative_hand": 26,
23
+ "real_r1_pro_sharpa_relative_eef_human": 26,
24
+ "real_r1_pro_sharpa_relative_eef_human_add_view": 26,
25
+ "real_r1_pro_sharpa_relative_eef_human_relative_hand": 26,
26
+ "real_r1_pro_sharpa_relative_eef_egodex": 26,
27
+ "real_r1_pro_sharpa_relative_eef_egodex_relative_hand": 26,
28
+ "real_r1_pro_sharpa_relative_eef_egodex_wrist_only": 26,
29
+ "real_r1_pro_sharpa_relative_eef_maxinsights": 26,
30
+ "real_r1_pro_sharpa_relative_eef_maxinsights_relative_hand": 26,
31
+ "real_r1_pro_sharpa_relative_eef_mecka": 26,
32
+ "real_r1_pro_sharpa_relative_eef_mecka_relative_hand": 26,
33
+ "real_g1_relative_eef_absolute_joints": 25,
34
+ "real_g1_relative_eef_absolute_joints_wrist_cam": 25,
35
+ "real_g1_relative_eef_relative_joints": 25,
36
+ "real_r1_pro_sharpa_relative_eef_relative_hand_relative_joint": 26,
37
+ "real_r1_pro_sharpa_relative_joint": 29,
38
+ "oxe_droid_relative_eef_relative_joint": 24,
39
+ "oxe_droid_relative_eef_relative_joint_swapped": 24,
40
+ "oxe_droid_relative_eef_relative_joint_upweight_z": 24,
41
+ "oxe_droid_relative_eef_relative_joint_upweight_z_swapped": 24,
42
+ "oxe_droid_relative_eef_relative_joint_3view": 24,
43
+ "oxe_droid_relative_eef_relative_joint_3view_swapped": 24,
44
+ "oxe_droid_relative_eef": 24,
45
+ "oxe_droid_joint_position_relative": 24,
46
+ "xdof_relative_eef_relative_joint": 27,
47
+ "xdof_relative_eef_relative_joint_subtask": 27,
48
+ "xdof_relative_eef": 27,
49
+ "xdof_relative_joint": 28,
50
+ "simpler_env_google": 0,
51
+ "simpler_env_widowx": 1,
52
+ "libero_sim": 2,
53
+ "droid_sim": 3,
54
+ "new_embodiment": 10,
55
+ "robocasa_gr1_tabletop": 10,
56
+ "unitree_g1_sonic": 11
57
+ }
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/experiment_cfg/conf.yaml ADDED
@@ -0,0 +1,251 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ load_config_path: null
2
+ model:
3
+ model_type: Gr00tN1d7
4
+ model_dtype: bfloat16
5
+ model_name: nvidia/Cosmos-Reason2-2B
6
+ backbone_model_type: qwen
7
+ model_revision: null
8
+ tune_top_llm_layers: 0
9
+ backbone_embedding_dim: 2048
10
+ tune_llm: true
11
+ tune_visual: true
12
+ select_layer: 12
13
+ reproject_vision: false
14
+ use_flash_attention: true
15
+ load_bf16: false
16
+ backbone_trainable_params_fp32: true
17
+ image_crop_size:
18
+ - 230
19
+ - 230
20
+ image_target_size:
21
+ - 256
22
+ - 256
23
+ shortest_image_edge: null
24
+ crop_fraction: null
25
+ random_rotation_angle: null
26
+ color_jitter_params: null
27
+ use_albumentations_transforms: true
28
+ letter_box_transform: false
29
+ extra_augmentation_config: null
30
+ formalize_language: true
31
+ apply_sincos_state_encoding: false
32
+ use_percentiles: true
33
+ use_relative_action: true
34
+ max_state_dim: 132
35
+ max_action_dim: 132
36
+ action_horizon: 40
37
+ strict_action_padding_mask: null
38
+ hidden_size: 1024
39
+ input_embedding_dim: 1536
40
+ state_history_length: 1
41
+ add_pos_embed: true
42
+ attn_dropout: 0.0
43
+ use_vlln: true
44
+ strict_vl_padding_mask: null
45
+ max_seq_len: 1024
46
+ use_alternate_vl_dit: true
47
+ attend_text_every_n_blocks: 2
48
+ diffusion_model_cfg:
49
+ positional_embeddings: null
50
+ num_layers: 16
51
+ num_attention_heads: 32
52
+ attention_head_dim: 48
53
+ norm_type: ada_norm
54
+ dropout: 0.0
55
+ final_dropout: true
56
+ output_dim: 1024
57
+ interleave_self_attention: true
58
+ num_inference_timesteps: 16
59
+ noise_beta_alpha: 1.5
60
+ noise_beta_beta: 1.0
61
+ noise_s: 0.999
62
+ num_timestep_buckets: 1000
63
+ tune_projector: true
64
+ tune_diffusion_model: true
65
+ tune_vlln: true
66
+ state_dropout_prob: 0.0
67
+ exclude_state: false
68
+ use_mean_std: false
69
+ max_num_embodiments: 32
70
+ data:
71
+ datasets:
72
+ - dataset_paths:
73
+ - /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/groot-data/a18cr2
74
+ embodiment_tag: new_embodiment
75
+ mix_ratio: 1.0
76
+ dataset_type: physical_embodiment
77
+ val_dataset_path: null
78
+ modality_configs:
79
+ new_embodiment:
80
+ video:
81
+ delta_indices:
82
+ - 0
83
+ modality_keys:
84
+ - head
85
+ - chest
86
+ sin_cos_embedding_keys: null
87
+ mean_std_embedding_keys: null
88
+ action_configs: null
89
+ state:
90
+ delta_indices:
91
+ - 0
92
+ modality_keys:
93
+ - left_leg
94
+ - right_leg
95
+ - waist
96
+ - left_arm
97
+ - right_arm
98
+ - head
99
+ - left_hand
100
+ - right_hand
101
+ - projected_gravity
102
+ sin_cos_embedding_keys: null
103
+ mean_std_embedding_keys: null
104
+ action_configs: null
105
+ action:
106
+ delta_indices:
107
+ - 0
108
+ - 1
109
+ - 2
110
+ - 3
111
+ - 4
112
+ - 5
113
+ - 6
114
+ - 7
115
+ - 8
116
+ - 9
117
+ - 10
118
+ - 11
119
+ - 12
120
+ - 13
121
+ - 14
122
+ - 15
123
+ - 16
124
+ - 17
125
+ - 18
126
+ - 19
127
+ - 20
128
+ - 21
129
+ - 22
130
+ - 23
131
+ - 24
132
+ - 25
133
+ - 26
134
+ - 27
135
+ - 28
136
+ - 29
137
+ - 30
138
+ - 31
139
+ modality_keys:
140
+ - body_target
141
+ - left_hand_target
142
+ - right_hand_target
143
+ sin_cos_embedding_keys: null
144
+ mean_std_embedding_keys: null
145
+ action_configs:
146
+ - rep: ABSOLUTE
147
+ type: NON_EEF
148
+ format: DEFAULT
149
+ state_key: null
150
+ - rep: ABSOLUTE
151
+ type: NON_EEF
152
+ format: DEFAULT
153
+ state_key: null
154
+ - rep: ABSOLUTE
155
+ type: NON_EEF
156
+ format: DEFAULT
157
+ state_key: null
158
+ language:
159
+ delta_indices:
160
+ - 0
161
+ modality_keys:
162
+ - annotation.human.task_description
163
+ sin_cos_embedding_keys: null
164
+ mean_std_embedding_keys: null
165
+ action_configs: null
166
+ download_cache: false
167
+ shard_size: 1024
168
+ episode_sampling_rate: 0.1
169
+ num_shards_per_epoch: 100000
170
+ ds_weights_alpha: null
171
+ override_pretraining_statistics: true
172
+ mode: single_turn
173
+ random_chop: 0.0
174
+ mock_dataset_mode: false
175
+ shuffle: true
176
+ seed: 42
177
+ multiprocessing_context: fork
178
+ allow_padding: false
179
+ subsample_ratio: 1.0
180
+ loader: fast
181
+ fast_loader:
182
+ decode_threads: 2
183
+ ffmpeg_threads: 1
184
+ shuffle_buffer: 0
185
+ min_fill: 0
186
+ image_crop_size:
187
+ - 244
188
+ - 244
189
+ image_target_size:
190
+ - 224
191
+ - 224
192
+ training:
193
+ output_dir: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/groot-runs-h32/a18cr2/production
194
+ experiment_name: null
195
+ max_steps: 4000
196
+ global_batch_size: 512
197
+ per_gpu_batch_size: null
198
+ gradient_accumulation_steps: 1
199
+ learning_rate: 0.0001
200
+ lr_scheduler_type: cosine
201
+ weight_decay: 1.0e-05
202
+ warmup_ratio: 0.05
203
+ warmup_steps: 0
204
+ max_grad_norm: 1.0
205
+ optim: adamw_torch
206
+ start_from_checkpoint: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/groot-n17-base
207
+ skip_weight_loading: false
208
+ tf32: true
209
+ fp16: false
210
+ bf16: true
211
+ eval_bf16: true
212
+ logging_steps: 10
213
+ save_steps: 500
214
+ save_total_limit: 5
215
+ save_vl_model: false
216
+ save_only_model: false
217
+ resume_from_checkpoint: false
218
+ upload_checkpoints: false
219
+ upload_every: 1000
220
+ upload_last_n_checkpoints: 5
221
+ max_concurrent_uploads: 2
222
+ eval_strategy: 'no'
223
+ eval_steps: 500
224
+ eval_set_split_ratio: 0.1
225
+ eval_batch_size: 2
226
+ save_best_eval_metric_name: ''
227
+ save_best_eval_metric_greater_is_better: true
228
+ deepspeed_stage: 2
229
+ gradient_checkpointing: false
230
+ transformers_trust_remote_code: true
231
+ transformers_local_files_only: false
232
+ transformers_cache_dir: null
233
+ transformers_access_token: null
234
+ use_ddp: false
235
+ ddp_bucket_cap_mb: 100
236
+ num_gpus: 16
237
+ dataloader_num_workers: 8
238
+ remove_unused_columns: false
239
+ use_wandb: false
240
+ wandb_project: finetune-gr00t-n1d7
241
+ enable_profiling: false
242
+ max_retries: 3
243
+ assert_loss_less_than: null
244
+ add_rl_callback: false
245
+ enable_open_loop_eval: false
246
+ open_loop_eval_traj_ids:
247
+ - 0
248
+ open_loop_eval_steps_per_traj: 100
249
+ open_loop_eval_plot_indices: null
250
+ max_steps: 4000
251
+ save_steps: 500
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/experiment_cfg/config.yaml ADDED
@@ -0,0 +1,249 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ load_config_path: null
2
+ model:
3
+ model_type: Gr00tN1d7
4
+ model_dtype: bfloat16
5
+ model_name: nvidia/Cosmos-Reason2-2B
6
+ backbone_model_type: qwen
7
+ model_revision: null
8
+ tune_top_llm_layers: 0
9
+ backbone_embedding_dim: 2048
10
+ tune_llm: true
11
+ tune_visual: true
12
+ select_layer: 12
13
+ reproject_vision: false
14
+ use_flash_attention: true
15
+ load_bf16: false
16
+ backbone_trainable_params_fp32: true
17
+ image_crop_size:
18
+ - 230
19
+ - 230
20
+ image_target_size:
21
+ - 256
22
+ - 256
23
+ shortest_image_edge: null
24
+ crop_fraction: null
25
+ random_rotation_angle: null
26
+ color_jitter_params: null
27
+ use_albumentations_transforms: true
28
+ letter_box_transform: false
29
+ extra_augmentation_config: null
30
+ formalize_language: true
31
+ apply_sincos_state_encoding: false
32
+ use_percentiles: true
33
+ use_relative_action: true
34
+ max_state_dim: 132
35
+ max_action_dim: 132
36
+ action_horizon: 40
37
+ strict_action_padding_mask: null
38
+ hidden_size: 1024
39
+ input_embedding_dim: 1536
40
+ state_history_length: 1
41
+ add_pos_embed: true
42
+ attn_dropout: 0.0
43
+ use_vlln: true
44
+ strict_vl_padding_mask: null
45
+ max_seq_len: 1024
46
+ use_alternate_vl_dit: true
47
+ attend_text_every_n_blocks: 2
48
+ diffusion_model_cfg:
49
+ positional_embeddings: null
50
+ num_layers: 16
51
+ num_attention_heads: 32
52
+ attention_head_dim: 48
53
+ norm_type: ada_norm
54
+ dropout: 0.0
55
+ final_dropout: true
56
+ output_dim: 1024
57
+ interleave_self_attention: true
58
+ num_inference_timesteps: 16
59
+ noise_beta_alpha: 1.5
60
+ noise_beta_beta: 1.0
61
+ noise_s: 0.999
62
+ num_timestep_buckets: 1000
63
+ tune_projector: true
64
+ tune_diffusion_model: true
65
+ tune_vlln: true
66
+ state_dropout_prob: 0.0
67
+ exclude_state: false
68
+ use_mean_std: false
69
+ max_num_embodiments: 32
70
+ data:
71
+ datasets:
72
+ - dataset_paths:
73
+ - /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/groot-data/a18cr2
74
+ embodiment_tag: new_embodiment
75
+ mix_ratio: 1.0
76
+ dataset_type: physical_embodiment
77
+ val_dataset_path: null
78
+ modality_configs:
79
+ new_embodiment:
80
+ video:
81
+ delta_indices:
82
+ - 0
83
+ modality_keys:
84
+ - head
85
+ - chest
86
+ sin_cos_embedding_keys: null
87
+ mean_std_embedding_keys: null
88
+ action_configs: null
89
+ state:
90
+ delta_indices:
91
+ - 0
92
+ modality_keys:
93
+ - left_leg
94
+ - right_leg
95
+ - waist
96
+ - left_arm
97
+ - right_arm
98
+ - head
99
+ - left_hand
100
+ - right_hand
101
+ - projected_gravity
102
+ sin_cos_embedding_keys: null
103
+ mean_std_embedding_keys: null
104
+ action_configs: null
105
+ action:
106
+ delta_indices:
107
+ - 0
108
+ - 1
109
+ - 2
110
+ - 3
111
+ - 4
112
+ - 5
113
+ - 6
114
+ - 7
115
+ - 8
116
+ - 9
117
+ - 10
118
+ - 11
119
+ - 12
120
+ - 13
121
+ - 14
122
+ - 15
123
+ - 16
124
+ - 17
125
+ - 18
126
+ - 19
127
+ - 20
128
+ - 21
129
+ - 22
130
+ - 23
131
+ - 24
132
+ - 25
133
+ - 26
134
+ - 27
135
+ - 28
136
+ - 29
137
+ - 30
138
+ - 31
139
+ modality_keys:
140
+ - body_target
141
+ - left_hand_target
142
+ - right_hand_target
143
+ sin_cos_embedding_keys: null
144
+ mean_std_embedding_keys: null
145
+ action_configs:
146
+ - rep: absolute
147
+ type: non_eef
148
+ format: default
149
+ state_key: null
150
+ - rep: absolute
151
+ type: non_eef
152
+ format: default
153
+ state_key: null
154
+ - rep: absolute
155
+ type: non_eef
156
+ format: default
157
+ state_key: null
158
+ language:
159
+ delta_indices:
160
+ - 0
161
+ modality_keys:
162
+ - annotation.human.task_description
163
+ sin_cos_embedding_keys: null
164
+ mean_std_embedding_keys: null
165
+ action_configs: null
166
+ download_cache: false
167
+ shard_size: 1024
168
+ episode_sampling_rate: 0.1
169
+ num_shards_per_epoch: 100000
170
+ ds_weights_alpha: null
171
+ override_pretraining_statistics: true
172
+ mode: single_turn
173
+ random_chop: 0.0
174
+ mock_dataset_mode: false
175
+ shuffle: true
176
+ seed: 42
177
+ multiprocessing_context: fork
178
+ allow_padding: false
179
+ subsample_ratio: 1.0
180
+ loader: fast
181
+ fast_loader:
182
+ decode_threads: 2
183
+ ffmpeg_threads: 1
184
+ shuffle_buffer: 0
185
+ min_fill: 0
186
+ image_crop_size:
187
+ - 244
188
+ - 244
189
+ image_target_size:
190
+ - 224
191
+ - 224
192
+ training:
193
+ output_dir: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/groot-runs-h32/a18cr2/production
194
+ experiment_name: null
195
+ max_steps: 4000
196
+ global_batch_size: 512
197
+ per_gpu_batch_size: null
198
+ gradient_accumulation_steps: 1
199
+ learning_rate: 0.0001
200
+ lr_scheduler_type: cosine
201
+ weight_decay: 1.0e-05
202
+ warmup_ratio: 0.05
203
+ warmup_steps: 0
204
+ max_grad_norm: 1.0
205
+ optim: adamw_torch
206
+ start_from_checkpoint: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/groot-n17-base
207
+ skip_weight_loading: false
208
+ tf32: true
209
+ fp16: false
210
+ bf16: true
211
+ eval_bf16: true
212
+ logging_steps: 10
213
+ save_steps: 500
214
+ save_total_limit: 5
215
+ save_vl_model: false
216
+ save_only_model: false
217
+ resume_from_checkpoint: false
218
+ upload_checkpoints: false
219
+ upload_every: 1000
220
+ upload_last_n_checkpoints: 5
221
+ max_concurrent_uploads: 2
222
+ eval_strategy: 'no'
223
+ eval_steps: 500
224
+ eval_set_split_ratio: 0.1
225
+ eval_batch_size: 2
226
+ save_best_eval_metric_name: ''
227
+ save_best_eval_metric_greater_is_better: true
228
+ deepspeed_stage: 2
229
+ gradient_checkpointing: false
230
+ transformers_trust_remote_code: true
231
+ transformers_local_files_only: false
232
+ transformers_cache_dir: null
233
+ transformers_access_token: null
234
+ use_ddp: false
235
+ ddp_bucket_cap_mb: 100
236
+ num_gpus: 16
237
+ dataloader_num_workers: 8
238
+ remove_unused_columns: false
239
+ use_wandb: false
240
+ wandb_project: finetune-gr00t-n1d7
241
+ enable_profiling: false
242
+ max_retries: 3
243
+ assert_loss_less_than: null
244
+ add_rl_callback: false
245
+ enable_open_loop_eval: false
246
+ open_loop_eval_traj_ids:
247
+ - 0
248
+ open_loop_eval_steps_per_traj: 100
249
+ open_loop_eval_plot_indices: null
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/experiment_cfg/dataset_statistics.json ADDED
@@ -0,0 +1,1095 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "new_embodiment": {
3
+ "state": {
4
+ "left_leg": {
5
+ "min": [
6
+ -0.12583781778812408,
7
+ -0.03233915567398071,
8
+ -0.12766100466251373,
9
+ 0.12355858832597733,
10
+ -0.06983537971973419,
11
+ -0.25010767579078674
12
+ ],
13
+ "max": [
14
+ 0.035650208592414856,
15
+ 0.10629415512084961,
16
+ 0.17557011544704437,
17
+ 0.32769498229026794,
18
+ 0.0809067115187645,
19
+ -0.03284743055701256
20
+ ],
21
+ "mean": [
22
+ -0.0272991843521595,
23
+ 0.03502087667584419,
24
+ 0.031001342460513115,
25
+ 0.16613589227199554,
26
+ 0.034384869039058685,
27
+ -0.12577565014362335
28
+ ],
29
+ "std": [
30
+ 0.03564170002937317,
31
+ 0.017918208613991737,
32
+ 0.038557928055524826,
33
+ 0.03397267311811447,
34
+ 0.022713234648108482,
35
+ 0.017287857830524445
36
+ ],
37
+ "q01": [
38
+ -0.09795097433030606,
39
+ 0.0007327128696488217,
40
+ -0.053566509149968626,
41
+ 0.13027908712625502,
42
+ -0.017819250598549843,
43
+ -0.15952911242842674
44
+ ],
45
+ "q99": [
46
+ 0.022375062648206902,
47
+ 0.0853965283930301,
48
+ 0.12469748824834821,
49
+ 0.2417573669552803,
50
+ 0.0676774867624044,
51
+ -0.0882675897330046
52
+ ]
53
+ },
54
+ "right_leg": {
55
+ "min": [
56
+ -0.12295906990766525,
57
+ -0.13229477405548096,
58
+ -0.1533408761024475,
59
+ 0.14615534245967865,
60
+ -0.03606459125876427,
61
+ -0.21855871379375458
62
+ ],
63
+ "max": [
64
+ 0.04188062623143196,
65
+ -1.1926664228667505e-05,
66
+ 0.22258582711219788,
67
+ 0.31865113973617554,
68
+ 0.03910355642437935,
69
+ -0.08754467964172363
70
+ ],
71
+ "mean": [
72
+ -0.022502576932311058,
73
+ -0.029816705733537674,
74
+ 0.015664661303162575,
75
+ 0.20424002408981323,
76
+ -0.0024386802688241005,
77
+ -0.15014690160751343
78
+ ],
79
+ "std": [
80
+ 0.032841239124536514,
81
+ 0.017189135774970055,
82
+ 0.05321863666176796,
83
+ 0.026313651353120804,
84
+ 0.011696949601173401,
85
+ 0.015430096536874771
86
+ ],
87
+ "q01": [
88
+ -0.0963956767320633,
89
+ -0.06397112980484962,
90
+ -0.09846653804183006,
91
+ 0.1504357412457466,
92
+ -0.027745265606790782,
93
+ -0.18931940764188768
94
+ ],
95
+ "q99": [
96
+ 0.02355051416903733,
97
+ -0.001509574134834111,
98
+ 0.14735825762152657,
99
+ 0.26482014119625086,
100
+ 0.02536030944436787,
101
+ -0.11168898373842247
102
+ ]
103
+ },
104
+ "waist": {
105
+ "min": [
106
+ -0.05953621491789818,
107
+ -0.11040446907281876,
108
+ -0.01577889733016491
109
+ ],
110
+ "max": [
111
+ 0.06784578412771225,
112
+ 0.030363162979483604,
113
+ 0.0996260941028595
114
+ ],
115
+ "mean": [
116
+ 0.00827261246740818,
117
+ -0.014616362750530243,
118
+ 0.03146738559007645
119
+ ],
120
+ "std": [
121
+ 0.01874351315200329,
122
+ 0.012871939688920975,
123
+ 0.02176288142800331
124
+ ],
125
+ "q01": [
126
+ -0.04488445371389389,
127
+ -0.05075038857758045,
128
+ -0.011339733125641942
129
+ ],
130
+ "q99": [
131
+ 0.04509926375001661,
132
+ 0.011364891706034533,
133
+ 0.0665438386797905
134
+ ]
135
+ },
136
+ "left_arm": {
137
+ "min": [
138
+ 0.05845891311764717,
139
+ 0.025421474128961563,
140
+ -0.3774237632751465,
141
+ 0.9427529573440552,
142
+ -0.1936495453119278,
143
+ -0.12422491610050201,
144
+ 0.00011441537935752422
145
+ ],
146
+ "max": [
147
+ 0.24996721744537354,
148
+ 0.14023271203041077,
149
+ 0.24500703811645508,
150
+ 1.1965162754058838,
151
+ 0.19641946256160736,
152
+ 0.30692338943481445,
153
+ 0.26775699853897095
154
+ ],
155
+ "mean": [
156
+ 0.14498291909694672,
157
+ 0.08285835385322571,
158
+ -0.0492297038435936,
159
+ 1.0528144836425781,
160
+ -0.00038467260310426354,
161
+ 0.11435367166996002,
162
+ 0.11888598650693893
163
+ ],
164
+ "std": [
165
+ 0.034364789724349976,
166
+ 0.01760222762823105,
167
+ 0.1176680251955986,
168
+ 0.04596450552344307,
169
+ 0.05561625212430954,
170
+ 0.09595136344432831,
171
+ 0.06368569284677505
172
+ ],
173
+ "q01": [
174
+ 0.08242766767740249,
175
+ 0.03789790570735931,
176
+ -0.3259463134407997,
177
+ 0.9694832515716553,
178
+ -0.14191590175032615,
179
+ -0.09522409550845623,
180
+ 0.020235377121716736
181
+ ],
182
+ "q99": [
183
+ 0.2127386112511158,
184
+ 0.12125491462647905,
185
+ 0.17541890650987624,
186
+ 1.1789501321315765,
187
+ 0.14806905061006542,
188
+ 0.28577067166566833,
189
+ 0.24580566853284835
190
+ ]
191
+ },
192
+ "right_arm": {
193
+ "min": [
194
+ -0.740268349647522,
195
+ -0.3696679174900055,
196
+ -0.6199629902839661,
197
+ -0.4936543107032776,
198
+ -0.09027265757322311,
199
+ -0.1852821260690689,
200
+ -0.9950985908508301
201
+ ],
202
+ "max": [
203
+ 0.6624637246131897,
204
+ 0.2643685042858124,
205
+ 0.4765041768550873,
206
+ 1.14264714717865,
207
+ 1.9407533407211304,
208
+ 0.4451823830604553,
209
+ 0.7307088375091553
210
+ ],
211
+ "mean": [
212
+ -0.21785785257816315,
213
+ -0.04526735842227936,
214
+ -0.06827006489038467,
215
+ 0.267977237701416,
216
+ 0.8951376080513,
217
+ 0.18325918912887573,
218
+ -0.21330057084560394
219
+ ],
220
+ "std": [
221
+ 0.35357531905174255,
222
+ 0.10000842809677124,
223
+ 0.2389989048242569,
224
+ 0.46367961168289185,
225
+ 0.5773417353630066,
226
+ 0.1706582009792328,
227
+ 0.34063631296157837
228
+ ],
229
+ "q01": [
230
+ -0.7036943805217742,
231
+ -0.27815148144960405,
232
+ -0.5371566104888916,
233
+ -0.43050183326005936,
234
+ -0.08400869272649288,
235
+ -0.12237099573016166,
236
+ -0.9130586802959442
237
+ ],
238
+ "q99": [
239
+ 0.5799816823005675,
240
+ 0.20428369268774946,
241
+ 0.4324539497494697,
242
+ 1.0882853162288666,
243
+ 1.6896514785289762,
244
+ 0.43551453173160554,
245
+ 0.6365156918764093
246
+ ]
247
+ },
248
+ "head": {
249
+ "min": [
250
+ 0.00010179156379308552,
251
+ -0.27579498291015625
252
+ ],
253
+ "max": [
254
+ 0.681303858757019,
255
+ 0.12099187821149826
256
+ ],
257
+ "mean": [
258
+ 0.5226783156394958,
259
+ -0.008672930300235748
260
+ ],
261
+ "std": [
262
+ 0.23635071516036987,
263
+ 0.10407927632331848
264
+ ],
265
+ "q01": [
266
+ 0.0054619968542829155,
267
+ -0.2461491858959198
268
+ ],
269
+ "q99": [
270
+ 0.6550031036138535,
271
+ 0.11107079438865185
272
+ ]
273
+ },
274
+ "left_hand": {
275
+ "min": [
276
+ 2.4212911853283003e-07,
277
+ 1.991089675357216e-07,
278
+ -2.484825699866633e-06,
279
+ 1.5511126605360914e-07,
280
+ -0.07711441069841385,
281
+ -0.0014401556691154838,
282
+ -0.001440082211047411,
283
+ 1.1005147371179191e-07,
284
+ -0.0014395025791600347,
285
+ -0.15826088190078735,
286
+ 4.014998822299276e-08,
287
+ 3.3918841069180417e-08,
288
+ -0.0014388926792889833,
289
+ 2.7533292978887403e-08,
290
+ -0.06904831528663635,
291
+ 6.039757582954053e-09,
292
+ 5.286712401186833e-09,
293
+ 2.0993500271515586e-08,
294
+ 4.5118535574317775e-09,
295
+ -0.04190514609217644,
296
+ 3.7183456313982788e-09,
297
+ -0.0002550320641603321
298
+ ],
299
+ "max": [
300
+ 0.001787454355508089,
301
+ 0.0017886327113956213,
302
+ 0.0001129926458816044,
303
+ 0.00179005297832191,
304
+ 0.07618467509746552,
305
+ -1.8496356233299593e-06,
306
+ -1.9247966065449873e-06,
307
+ 0.0017915134085342288,
308
+ -1.7791824120649835e-06,
309
+ -0.00012538964801933616,
310
+ 0.000348125264281407,
311
+ 0.00034833172685466707,
312
+ -1.6332217001036042e-06,
313
+ 0.0003485499764792621,
314
+ 0.0031303067225962877,
315
+ 6.412907532649115e-05,
316
+ 6.415732786990702e-05,
317
+ 0.0003487732319626957,
318
+ 6.418681005015969e-05,
319
+ 0.04411711171269417,
320
+ 6.421708530979231e-05,
321
+ 0.00032537104561924934
322
+ ],
323
+ "mean": [
324
+ 0.0011903581907972693,
325
+ 0.001190043636597693,
326
+ 1.3743877389060799e-05,
327
+ 0.0011899300152435899,
328
+ 0.009631289169192314,
329
+ -0.0008670862880535424,
330
+ -0.0008672947296872735,
331
+ 0.001189946779049933,
332
+ -0.0008662287145853043,
333
+ -0.09532451629638672,
334
+ 0.00023694503761362284,
335
+ 0.00023693856201134622,
336
+ -0.0008651672396808863,
337
+ 0.0002369351132074371,
338
+ -0.010153159499168396,
339
+ 4.455091038835235e-05,
340
+ 4.455030648387037e-05,
341
+ 0.0002369288558838889,
342
+ 4.454981899471022e-05,
343
+ 0.0005994683015160263,
344
+ 4.4548927689902484e-05,
345
+ 2.5255261789425276e-05
346
+ ],
347
+ "std": [
348
+ 0.0003766888694372028,
349
+ 0.0003766732115764171,
350
+ 2.0943400159012526e-05,
351
+ 0.00037665216950699687,
352
+ 0.024747461080551147,
353
+ 0.00038130549364723265,
354
+ 0.0003809256595559418,
355
+ 0.00037665446870960295,
356
+ 0.000381362100597471,
357
+ 0.01209488976746794,
358
+ 7.698861736571416e-05,
359
+ 7.698720582993701e-05,
360
+ 0.000381679245037958,
361
+ 7.698764238739386e-05,
362
+ 0.013209239579737186,
363
+ 1.49271918417071e-05,
364
+ 1.4926949006621726e-05,
365
+ 7.698717672610655e-05,
366
+ 1.4926786207070109e-05,
367
+ 0.008573796600103378,
368
+ 1.49265251820907e-05,
369
+ 3.794630538322963e-05
370
+ ],
371
+ "q01": [
372
+ 5.6644346368557305e-05,
373
+ 5.6655935477465394e-05,
374
+ -1.0230358566332143e-06,
375
+ 5.670339669450186e-05,
376
+ -0.05207991287112236,
377
+ -0.0013900775543879717,
378
+ -0.0013900037901476026,
379
+ 5.677977540472057e-05,
380
+ -0.001389486389234662,
381
+ -0.12836721882224084,
382
+ 1.0760352224679082e-05,
383
+ 1.0762099882413168e-05,
384
+ -0.0013888693251647055,
385
+ 1.0771808256322401e-05,
386
+ -0.05400261629372835,
387
+ 1.9243952783654093e-06,
388
+ 1.9247009936407267e-06,
389
+ 1.078338749721297e-05,
390
+ 1.9262132514086262e-06,
391
+ -0.02179745353758335,
392
+ 1.927753485233552e-06,
393
+ -4.8397314858448224e-05
394
+ ],
395
+ "q99": [
396
+ 0.0016224077937658876,
397
+ 0.0016217380098532885,
398
+ 7.156058636610374e-05,
399
+ 0.0016222188377287236,
400
+ 0.04415374629199504,
401
+ -3.119549841358094e-05,
402
+ -3.223132720449939e-05,
403
+ 0.0016219218051992358,
404
+ -3.022737266292097e-05,
405
+ -0.07546421341598034,
406
+ 0.00032220381370279933,
407
+ 0.000322212442406453,
408
+ -2.8710181559290505e-05,
409
+ 0.00032226746581727624,
410
+ 0.0029770207405090316,
411
+ 6.057487513317027e-05,
412
+ 6.0567598229681604e-05,
413
+ 0.00032231434131972474,
414
+ 6.056960286514366e-05,
415
+ 0.026138318255543698,
416
+ 6.0573126065719385e-05,
417
+ 0.00011700723756803194
418
+ ]
419
+ },
420
+ "right_hand": {
421
+ "min": [
422
+ -0.0012902324087917805,
423
+ -0.011018202640116215,
424
+ -1.4729690747117274e-06,
425
+ -0.008869198150932789,
426
+ -0.0010071652941405773,
427
+ -0.07318796217441559,
428
+ -0.14104261994361877,
429
+ -0.006404141429811716,
430
+ -0.012594389729201794,
431
+ -0.2601270377635956,
432
+ -0.005105330143123865,
433
+ -0.0029860667418688536,
434
+ -0.000577878556214273,
435
+ -0.004302919376641512,
436
+ -0.003525838255882263,
437
+ -0.0016964594833552837,
438
+ -0.0017751255072653294,
439
+ -0.004758988507091999,
440
+ 1.0182914067513593e-08,
441
+ -0.01523496862500906,
442
+ -0.0019727651961147785,
443
+ -0.0021971184760332108
444
+ ],
445
+ "max": [
446
+ 0.38441139459609985,
447
+ 0.7239343523979187,
448
+ 0.041639670729637146,
449
+ 0.6621608734130859,
450
+ 0.7537572383880615,
451
+ 0.01115428563207388,
452
+ 0.003058104310184717,
453
+ 0.5719240307807922,
454
+ 0.0017499510431662202,
455
+ -7.032280700514093e-05,
456
+ 1.7279481887817383,
457
+ 1.6268699169158936,
458
+ 0.14201073348522186,
459
+ 1.5546116828918457,
460
+ 0.5081002116203308,
461
+ 1.247889518737793,
462
+ 1.182234525680542,
463
+ 1.7133269309997559,
464
+ 0.27388203144073486,
465
+ 0.014611090533435345,
466
+ 0.9894230365753174,
467
+ 0.8894627094268799
468
+ ],
469
+ "mean": [
470
+ 0.2006378173828125,
471
+ 0.39451727271080017,
472
+ 0.021642182022333145,
473
+ 0.3775778114795685,
474
+ 0.4183473289012909,
475
+ -0.02677745372056961,
476
+ -0.07673882693052292,
477
+ 0.3158021867275238,
478
+ -0.004371536895632744,
479
+ -0.17732049524784088,
480
+ 0.9738762974739075,
481
+ 0.9352828860282898,
482
+ 0.08735551685094833,
483
+ 0.9004462361335754,
484
+ 0.26746442914009094,
485
+ 0.7129588723182678,
486
+ 0.6785854697227478,
487
+ 0.986158013343811,
488
+ 0.15895037353038788,
489
+ -0.003946554847061634,
490
+ 0.5721363425254822,
491
+ 0.4799375832080841
492
+ ],
493
+ "std": [
494
+ 0.16025732457637787,
495
+ 0.31994420289993286,
496
+ 0.014540785923600197,
497
+ 0.3069762587547302,
498
+ 0.34989607334136963,
499
+ 0.021697280928492546,
500
+ 0.05449501797556877,
501
+ 0.25693169236183167,
502
+ 0.003981089685112238,
503
+ 0.07617250829935074,
504
+ 0.7627589106559753,
505
+ 0.7256377935409546,
506
+ 0.06025630608201027,
507
+ 0.7178282141685486,
508
+ 0.2390982061624527,
509
+ 0.5618950724601746,
510
+ 0.5340582728385925,
511
+ 0.7808806300163269,
512
+ 0.12609755992889404,
513
+ 0.005126067902892828,
514
+ 0.45124730467796326,
515
+ 0.42998701333999634
516
+ ],
517
+ "q01": [
518
+ -0.0008357079606503248,
519
+ -0.008844400178641082,
520
+ 1.720549362289603e-05,
521
+ -0.007531417375430465,
522
+ 0.009923810744658113,
523
+ -0.06775790706276894,
524
+ -0.1352728369832039,
525
+ -0.005800717449747026,
526
+ -0.010245489543303848,
527
+ -0.2565269756317139,
528
+ 1.0928990704996976e-05,
529
+ 8.868164959494607e-06,
530
+ -0.0005466874601552263,
531
+ 1.0923624004135492e-05,
532
+ -0.003336994650308043,
533
+ 1.5674098028739536e-06,
534
+ 1.5685043285884604e-06,
535
+ 1.0920326503764955e-05,
536
+ 1.9563896034924257e-06,
537
+ -0.012814011462032794,
538
+ 1.6161931728220225e-06,
539
+ 2.7791540901489496e-06
540
+ ],
541
+ "q99": [
542
+ 0.37896424770355225,
543
+ 0.7207409799098968,
544
+ 0.04077165637165308,
545
+ 0.6620168089866638,
546
+ 0.7518837815523147,
547
+ 0.005935645233839748,
548
+ 0.0029098605737090106,
549
+ 0.5708987444639206,
550
+ 0.0014530597848352045,
551
+ -0.05925159651786089,
552
+ 1.7104909121990204,
553
+ 1.6180911695957183,
554
+ 0.13977389827370643,
555
+ 1.5544607639312744,
556
+ 0.503916236758232,
557
+ 1.2390973663330078,
558
+ 1.1762168061733245,
559
+ 1.711379610300064,
560
+ 0.27380754083395004,
561
+ 0.009059532321989505,
562
+ 0.9886172729730606,
563
+ 0.8870919644832611
564
+ ]
565
+ },
566
+ "projected_gravity": {
567
+ "min": [
568
+ -0.08454640954732895,
569
+ -0.09836447238922119,
570
+ -1.0
571
+ ],
572
+ "max": [
573
+ 0.00796288251876831,
574
+ 0.036943208426237106,
575
+ -0.9915630221366882
576
+ ],
577
+ "mean": [
578
+ -0.02378002367913723,
579
+ -0.010027356445789337,
580
+ -0.9998267889022827
581
+ ],
582
+ "std": [
583
+ 0.014949320815503597,
584
+ 0.009925544261932373,
585
+ 0.0005525829037781774
586
+ ],
587
+ "q01": [
588
+ -0.0513697436824441,
589
+ -0.03479166690260172,
590
+ -0.9999969601631165
591
+ ],
592
+ "q99": [
593
+ 0.003534715329296886,
594
+ 0.017252918723970646,
595
+ -0.9981654179096222
596
+ ]
597
+ }
598
+ },
599
+ "action": {
600
+ "body_target": {
601
+ "min": [
602
+ -0.1784663051366806,
603
+ -0.13318756222724915,
604
+ -0.1307600885629654,
605
+ 0.0,
606
+ -0.05743294209241867,
607
+ -0.3024138808250427,
608
+ -0.12985824048519135,
609
+ -0.266737699508667,
610
+ -0.15798582136631012,
611
+ 0.0,
612
+ -0.0613182857632637,
613
+ -0.4162753224372864,
614
+ -0.06568841636180878,
615
+ -0.21089573204517365,
616
+ -0.046035848557949066,
617
+ 0.0,
618
+ 0.0,
619
+ -0.40408286452293396,
620
+ 0.0,
621
+ -0.19912302494049072,
622
+ -0.156949982047081,
623
+ 0.0,
624
+ -0.7415577173233032,
625
+ -0.37339621782302856,
626
+ -0.6357685327529907,
627
+ -0.4724564254283905,
628
+ -0.1029481366276741,
629
+ -0.1933467984199524,
630
+ -0.99897301197052,
631
+ 0.0,
632
+ -0.2813774645328522
633
+ ],
634
+ "max": [
635
+ 0.041423577815294266,
636
+ 0.24960044026374817,
637
+ 0.17349489033222198,
638
+ 0.3818897604942322,
639
+ 0.09517528861761093,
640
+ 0.33370673656463623,
641
+ 0.08997327834367752,
642
+ 0.1119127869606018,
643
+ 0.2282232642173767,
644
+ 0.3609016537666321,
645
+ 0.020472919568419456,
646
+ 0.19194680452346802,
647
+ 0.0703074038028717,
648
+ 0.03390442579984665,
649
+ 0.16752633452415466,
650
+ 0.2617945075035095,
651
+ 0.21634480357170105,
652
+ 0.2762379050254822,
653
+ 1.1804041862487793,
654
+ 0.20910896360874176,
655
+ 0.33972981572151184,
656
+ 0.28348755836486816,
657
+ 0.5959668159484863,
658
+ 0.2778860032558441,
659
+ 0.4788244366645813,
660
+ 1.1417303085327148,
661
+ 1.9401861429214478,
662
+ 0.4363322854042053,
663
+ 0.737479031085968,
664
+ 0.6173598170280457,
665
+ 0.12206494063138962
666
+ ],
667
+ "mean": [
668
+ -0.03915031999349594,
669
+ 0.041799839586019516,
670
+ 0.03114175610244274,
671
+ 0.1608261913061142,
672
+ 0.05141807720065117,
673
+ -0.15212038159370422,
674
+ -0.00847386009991169,
675
+ -0.024333275854587555,
676
+ 0.016389140859246254,
677
+ 0.22511744499206543,
678
+ -0.02549057826399803,
679
+ -0.032981500029563904,
680
+ 0.009544417262077332,
681
+ -0.05076265707612038,
682
+ 0.03523816913366318,
683
+ 0.1433994024991989,
684
+ 0.11653806269168854,
685
+ -0.04866323620080948,
686
+ 0.9892690181732178,
687
+ 0.004534699022769928,
688
+ 0.09968917816877365,
689
+ 0.1222209483385086,
690
+ -0.21160988509655,
691
+ -0.0468902513384819,
692
+ -0.06461045891046524,
693
+ 0.26041075587272644,
694
+ 0.8931002020835876,
695
+ 0.18262670934200287,
696
+ -0.2106882780790329,
697
+ 0.4847927391529083,
698
+ -0.008713728748261929
699
+ ],
700
+ "std": [
701
+ 0.04093769192695618,
702
+ 0.03702395036816597,
703
+ 0.03900440037250519,
704
+ 0.0471014603972435,
705
+ 0.026934955269098282,
706
+ 0.05626444146037102,
707
+ 0.04267170652747154,
708
+ 0.03333482891321182,
709
+ 0.05455751717090607,
710
+ 0.03382829576730728,
711
+ 0.013083825819194317,
712
+ 0.07339824736118317,
713
+ 0.01564890705049038,
714
+ 0.03649556636810303,
715
+ 0.03141408786177635,
716
+ 0.05130252242088318,
717
+ 0.018622595816850662,
718
+ 0.1216823160648346,
719
+ 0.08364145457744598,
720
+ 0.05608460679650307,
721
+ 0.10647325217723846,
722
+ 0.0722256749868393,
723
+ 0.35748088359832764,
724
+ 0.09978532046079636,
725
+ 0.23952510952949524,
726
+ 0.4489462971687317,
727
+ 0.5780932307243347,
728
+ 0.17561344802379608,
729
+ 0.34522509574890137,
730
+ 0.21979530155658722,
731
+ 0.10528187453746796
732
+ ],
733
+ "q01": [
734
+ -0.132551628947258,
735
+ -0.024366050101816653,
736
+ -0.05374816991388798,
737
+ 0.09542277917265891,
738
+ -0.0036514422902837393,
739
+ -0.23983699575066567,
740
+ -0.10000019520521164,
741
+ -0.1274168734252453,
742
+ -0.10001292392611504,
743
+ 0.15025393679738044,
744
+ -0.04804414350539446,
745
+ -0.14999812841415405,
746
+ -0.03626127932220698,
747
+ -0.14458654195070267,
748
+ -0.015955421850085257,
749
+ 0.06150256689637899,
750
+ 0.07787629395723343,
751
+ -0.338915196955204,
752
+ 0.8881164407730102,
753
+ -0.13743558526039124,
754
+ -0.11451421245932579,
755
+ 1.830736131523736e-05,
756
+ -0.7038814079761505,
757
+ -0.2810768815875053,
758
+ -0.5427191376686096,
759
+ -0.40825496792793275,
760
+ -0.08908483155071735,
761
+ -0.1366129769384861,
762
+ -0.9141404831409454,
763
+ 0.0,
764
+ -0.248400519490242
765
+ ],
766
+ "q99": [
767
+ 0.022075342386960975,
768
+ 0.16839663177728642,
769
+ 0.12676575616002073,
770
+ 0.24997984781861304,
771
+ 0.08028831973671913,
772
+ 0.04339780144393434,
773
+ 0.051452237330377064,
774
+ 0.03742872074246399,
775
+ 0.14983677625656067,
776
+ 0.30238067269325253,
777
+ 1.084890353922674e-05,
778
+ 0.10649204172193996,
779
+ 0.044122488945722474,
780
+ 0.00013618171229609283,
781
+ 0.1245535047352313,
782
+ 0.24998776614665985,
783
+ 0.15528049811720845,
784
+ 0.1868483006954193,
785
+ 1.1411199951171875,
786
+ 0.1543787719309326,
787
+ 0.3053547310829161,
788
+ 0.25633819222450255,
789
+ 0.5314181041717528,
790
+ 0.1988790354132652,
791
+ 0.4322459593415258,
792
+ 1.082061173915863,
793
+ 1.6835209715366362,
794
+ 0.4363322854042053,
795
+ 0.63710262954235,
796
+ 0.6076847434043884,
797
+ 0.11178897745907303
798
+ ]
799
+ },
800
+ "left_hand_target": {
801
+ "min": [
802
+ 0.0,
803
+ 0.0,
804
+ 0.0,
805
+ 0.0,
806
+ 0.0,
807
+ 0.0,
808
+ 0.0,
809
+ 0.0,
810
+ 0.0,
811
+ 0.0,
812
+ 0.0,
813
+ 0.0,
814
+ 0.0,
815
+ 0.0,
816
+ 0.0,
817
+ 0.0,
818
+ 0.0,
819
+ 0.0,
820
+ 0.0,
821
+ 0.0,
822
+ 0.0,
823
+ 0.0
824
+ ],
825
+ "max": [
826
+ 0.0,
827
+ 0.0,
828
+ 0.0,
829
+ 0.0,
830
+ 0.0,
831
+ 0.0,
832
+ 0.0,
833
+ 0.0,
834
+ 0.0,
835
+ 0.0,
836
+ 0.0,
837
+ 0.0,
838
+ 0.0,
839
+ 0.0,
840
+ 0.0,
841
+ 0.0,
842
+ 0.0,
843
+ 0.0,
844
+ 0.0,
845
+ 0.0,
846
+ 0.0,
847
+ 0.0
848
+ ],
849
+ "mean": [
850
+ 0.0,
851
+ 0.0,
852
+ 0.0,
853
+ 0.0,
854
+ 0.0,
855
+ 0.0,
856
+ 0.0,
857
+ 0.0,
858
+ 0.0,
859
+ 0.0,
860
+ 0.0,
861
+ 0.0,
862
+ 0.0,
863
+ 0.0,
864
+ 0.0,
865
+ 0.0,
866
+ 0.0,
867
+ 0.0,
868
+ 0.0,
869
+ 0.0,
870
+ 0.0,
871
+ 0.0
872
+ ],
873
+ "std": [
874
+ 0.0,
875
+ 0.0,
876
+ 0.0,
877
+ 0.0,
878
+ 0.0,
879
+ 0.0,
880
+ 0.0,
881
+ 0.0,
882
+ 0.0,
883
+ 0.0,
884
+ 0.0,
885
+ 0.0,
886
+ 0.0,
887
+ 0.0,
888
+ 0.0,
889
+ 0.0,
890
+ 0.0,
891
+ 0.0,
892
+ 0.0,
893
+ 0.0,
894
+ 0.0,
895
+ 0.0
896
+ ],
897
+ "q01": [
898
+ 0.0,
899
+ 0.0,
900
+ 0.0,
901
+ 0.0,
902
+ 0.0,
903
+ 0.0,
904
+ 0.0,
905
+ 0.0,
906
+ 0.0,
907
+ 0.0,
908
+ 0.0,
909
+ 0.0,
910
+ 0.0,
911
+ 0.0,
912
+ 0.0,
913
+ 0.0,
914
+ 0.0,
915
+ 0.0,
916
+ 0.0,
917
+ 0.0,
918
+ 0.0,
919
+ 0.0
920
+ ],
921
+ "q99": [
922
+ 0.0,
923
+ 0.0,
924
+ 0.0,
925
+ 0.0,
926
+ 0.0,
927
+ 0.0,
928
+ 0.0,
929
+ 0.0,
930
+ 0.0,
931
+ 0.0,
932
+ 0.0,
933
+ 0.0,
934
+ 0.0,
935
+ 0.0,
936
+ 0.0,
937
+ 0.0,
938
+ 0.0,
939
+ 0.0,
940
+ 0.0,
941
+ 0.0,
942
+ 0.0,
943
+ 0.0
944
+ ]
945
+ },
946
+ "right_hand_target": {
947
+ "min": [
948
+ 0.0,
949
+ 0.0,
950
+ 0.0,
951
+ 0.0,
952
+ 0.0,
953
+ 0.0,
954
+ -0.2870999872684479,
955
+ 0.0,
956
+ -0.01590000092983246,
957
+ -0.17630000412464142,
958
+ 0.0,
959
+ 0.0,
960
+ 0.0,
961
+ 0.0,
962
+ 0.0,
963
+ 0.0,
964
+ 0.0,
965
+ 0.0,
966
+ 0.0,
967
+ -0.006200000178068876,
968
+ 0.0,
969
+ 0.0
970
+ ],
971
+ "max": [
972
+ 0.3626999855041504,
973
+ 0.777400016784668,
974
+ 0.04019999876618385,
975
+ 0.6610000133514404,
976
+ 0.7699999809265137,
977
+ 0.07270000129938126,
978
+ 0.0,
979
+ 0.5733000040054321,
980
+ 0.0,
981
+ 0.0,
982
+ 1.745300054550171,
983
+ 1.745300054550171,
984
+ 0.12300000339746475,
985
+ 1.5556999444961548,
986
+ 0.5410000085830688,
987
+ 1.2592999935150146,
988
+ 1.2130999565124512,
989
+ 1.745300054550171,
990
+ 0.27469998598098755,
991
+ 0.0,
992
+ 1.0023000240325928,
993
+ 0.8996000289916992
994
+ ],
995
+ "mean": [
996
+ 0.20521487295627594,
997
+ 0.4470158815383911,
998
+ 0.02411259338259697,
999
+ 0.3790503144264221,
1000
+ 0.42076605558395386,
1001
+ 0.04735369235277176,
1002
+ -0.18692949414253235,
1003
+ 0.32813677191734314,
1004
+ -0.010356616228818893,
1005
+ -0.11476758867502213,
1006
+ 1.0204970836639404,
1007
+ 1.0230900049209595,
1008
+ 0.08010388910770416,
1009
+ 0.9013310074806213,
1010
+ 0.2932063639163971,
1011
+ 0.7347167730331421,
1012
+ 0.7081287503242493,
1013
+ 1.0148118734359741,
1014
+ 0.1608325093984604,
1015
+ -0.004035439807921648,
1016
+ 0.5850551128387451,
1017
+ 0.49043309688568115
1018
+ ],
1019
+ "std": [
1020
+ 0.16896863281726837,
1021
+ 0.36067304015159607,
1022
+ 0.01834881491959095,
1023
+ 0.30681031942367554,
1024
+ 0.3792387545108795,
1025
+ 0.03370736166834831,
1026
+ 0.1331154704093933,
1027
+ 0.26620402932167053,
1028
+ 0.00737454229965806,
1029
+ 0.08175445348024368,
1030
+ 0.8066027164459229,
1031
+ 0.8055548071861267,
1032
+ 0.05705483257770538,
1033
+ 0.7204705476760864,
1034
+ 0.2650739252567291,
1035
+ 0.5830318927764893,
1036
+ 0.561620831489563,
1037
+ 0.8082879781723022,
1038
+ 0.12688195705413818,
1039
+ 0.002874528057873249,
1040
+ 0.46396660804748535,
1041
+ 0.4402594268321991
1042
+ ],
1043
+ "q01": [
1044
+ 0.0,
1045
+ 0.0,
1046
+ 0.0,
1047
+ 0.0,
1048
+ 0.0,
1049
+ 0.0,
1050
+ -0.2870999872684479,
1051
+ 0.0,
1052
+ -0.01590000092983246,
1053
+ -0.17630000412464142,
1054
+ 0.0,
1055
+ 0.0,
1056
+ 0.0,
1057
+ 0.0,
1058
+ 0.0,
1059
+ 0.0,
1060
+ 0.0,
1061
+ 0.0,
1062
+ 0.0,
1063
+ -0.006200000178068876,
1064
+ 0.0,
1065
+ 0.0
1066
+ ],
1067
+ "q99": [
1068
+ 0.3626999855041504,
1069
+ 0.777400016784668,
1070
+ 0.04019999876618385,
1071
+ 0.6610000133514404,
1072
+ 0.7699999809265137,
1073
+ 0.07270000129938126,
1074
+ 0.0,
1075
+ 0.5733000040054321,
1076
+ 0.0,
1077
+ 0.0,
1078
+ 1.745300054550171,
1079
+ 1.745300054550171,
1080
+ 0.12300000339746475,
1081
+ 1.5556999444961548,
1082
+ 0.5410000085830688,
1083
+ 1.2592999935150146,
1084
+ 1.2130999565124512,
1085
+ 1.745300054550171,
1086
+ 0.27469998598098755,
1087
+ 0.0,
1088
+ 1.0023000240325928,
1089
+ 0.8996000289916992
1090
+ ]
1091
+ }
1092
+ },
1093
+ "relative_action": {}
1094
+ }
1095
+ }
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/experiment_cfg/final_model_config.json ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model_type": "Gr00tN1d7",
3
+ "model_dtype": "bfloat16",
4
+ "model_name": "nvidia/Cosmos-Reason2-2B",
5
+ "backbone_model_type": "qwen",
6
+ "model_revision": null,
7
+ "tune_top_llm_layers": 0,
8
+ "backbone_embedding_dim": 2048,
9
+ "tune_llm": true,
10
+ "tune_visual": true,
11
+ "select_layer": 16,
12
+ "reproject_vision": false,
13
+ "use_flash_attention": true,
14
+ "load_bf16": false,
15
+ "backbone_trainable_params_fp32": true,
16
+ "letter_box_transform": false,
17
+ "extra_augmentation_config": null,
18
+ "apply_sincos_state_encoding": false,
19
+ "use_percentiles": true,
20
+ "use_relative_action": false,
21
+ "max_state_dim": 132,
22
+ "max_action_dim": 132,
23
+ "action_horizon": 40,
24
+ "strict_action_padding_mask": false,
25
+ "hidden_size": 1024,
26
+ "input_embedding_dim": 1536,
27
+ "state_history_length": 1,
28
+ "add_pos_embed": true,
29
+ "attn_dropout": 0.2,
30
+ "use_vlln": true,
31
+ "strict_vl_padding_mask": false,
32
+ "max_seq_len": 1024,
33
+ "use_alternate_vl_dit": true,
34
+ "attend_text_every_n_blocks": 2,
35
+ "diffusion_model_cfg": {
36
+ "attention_head_dim": 48,
37
+ "dropout": 0.2,
38
+ "final_dropout": true,
39
+ "interleave_self_attention": true,
40
+ "norm_type": "ada_norm",
41
+ "num_attention_heads": 32,
42
+ "num_layers": 32,
43
+ "output_dim": 1024,
44
+ "positional_embeddings": null
45
+ },
46
+ "num_inference_timesteps": 4,
47
+ "noise_beta_alpha": 1.5,
48
+ "noise_beta_beta": 1.0,
49
+ "noise_s": 0.999,
50
+ "num_timestep_buckets": 1000,
51
+ "tune_projector": true,
52
+ "tune_diffusion_model": true,
53
+ "tune_vlln": true,
54
+ "state_dropout_prob": 0.0,
55
+ "exclude_state": false,
56
+ "use_mean_std": false,
57
+ "max_num_embodiments": 32
58
+ }
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/experiment_cfg/final_processor_config.json ADDED
The diff for this file is too large to render. See raw diff
 
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/loss_curve.csv ADDED
@@ -0,0 +1,401 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ step,loss,learning_rate,grad_norm
2
+ 10,1.3905,4.5000000000000003e-07,0.24912892282009125
3
+ 20,1.3802,9.500000000000001e-07,0.394233375787735
4
+ 30,1.3083,1.45e-06,1.497800588607788
5
+ 40,1.2031,1.9500000000000004e-06,1.1424145698547363
6
+ 50,1.1098,2.4500000000000003e-06,1.2752643823623657
7
+ 60,1.079,2.95e-06,1.0560007095336914
8
+ 70,1.0589,3.45e-06,0.720940351486206
9
+ 80,1.043,3.95e-06,0.4508174955844879
10
+ 90,1.0379,4.450000000000001e-06,0.4460732042789459
11
+ 100,1.0347,4.95e-06,0.2577953338623047
12
+ 110,1.032,5.450000000000001e-06,0.5169486403465271
13
+ 120,1.0248,5.950000000000001e-06,0.44850045442581177
14
+ 130,1.0148,6.450000000000001e-06,0.5352460145950317
15
+ 140,0.9866,6.95e-06,0.738865315914154
16
+ 150,0.9231,7.450000000000001e-06,0.6277132034301758
17
+ 160,0.8618,7.950000000000002e-06,0.5653408169746399
18
+ 170,0.8111,8.45e-06,0.5750411152839661
19
+ 180,0.7641,8.95e-06,0.7955657839775085
20
+ 190,0.7216,9.450000000000001e-06,0.8227806091308594
21
+ 200,0.6926,9.950000000000001e-06,0.6388406753540039
22
+ 210,0.6776,9.999861593790126e-06,0.6702894568443298
23
+ 220,0.6523,9.999383162408303e-06,0.6838157176971436
24
+ 230,0.6323,9.998563029828258e-06,0.6812769770622253
25
+ 240,0.6057,9.997401252104962e-06,0.6650444865226746
26
+ 250,0.5821,9.995897908644378e-06,1.195233702659607
27
+ 260,0.5615,9.994053102198035e-06,0.8008884191513062
28
+ 270,0.5517,9.991866958856003e-06,0.744047999382019
29
+ 280,0.5187,9.989339628038277e-06,0.7080017328262329
30
+ 290,0.5128,9.98647128248456e-06,0.7555558085441589
31
+ 300,0.4822,9.983262118242461e-06,0.7810933589935303
32
+ 310,0.4704,9.979712354654092e-06,0.8074049949645996
33
+ 320,0.4418,9.97582223434108e-06,0.8137549757957458
34
+ 330,0.4221,9.971592023187979e-06,0.8988544940948486
35
+ 340,0.3992,9.967022010324105e-06,0.8715605139732361
36
+ 350,0.3715,9.962112508103766e-06,0.9414624571800232
37
+ 360,0.3612,9.956863852084914e-06,1.1080490350723267
38
+ 370,0.3463,9.95127640100622e-06,0.9747459292411804
39
+ 380,0.3182,9.945350536762544e-06,0.8970965147018433
40
+ 390,0.304,9.93908666437883e-06,1.0791109800338745
41
+ 400,0.2832,9.932485211982437e-06,1.1977298259735107
42
+ 410,0.272,9.92554663077387e-06,1.120720624923706
43
+ 420,0.2509,9.918271394995936e-06,1.2665119171142578
44
+ 430,0.2406,9.910660001901335e-06,1.0502629280090332
45
+ 440,0.2187,9.902712971718676e-06,1.1278620958328247
46
+ 450,0.2001,9.894430847616916e-06,0.9573274254798889
47
+ 460,0.1847,9.885814195668232e-06,0.9732072949409485
48
+ 470,0.1746,9.876863604809345e-06,1.0263832807540894
49
+ 480,0.1565,9.867579686801245e-06,1.1131765842437744
50
+ 490,0.143,9.8579630761874e-06,1.0001071691513062
51
+ 500,0.1289,9.848014430250368e-06,1.2044321298599243
52
+ 510,0.1252,9.837734428966885e-06,1.260959267616272
53
+ 520,0.1066,9.827123774961383e-06,1.1621663570404053
54
+ 530,0.1068,9.81618319345797e-06,1.0683119297027588
55
+ 540,0.0993,9.804913432230856e-06,1.4515645503997803
56
+ 550,0.0915,9.793315261553252e-06,0.8918207287788391
57
+ 560,0.0821,9.781389474144718e-06,0.9777098298072815
58
+ 570,0.0792,9.769136885116981e-06,0.9528394937515259
59
+ 580,0.0809,9.756558331918227e-06,0.9681083559989929
60
+ 590,0.0822,9.743654674275855e-06,0.7851548790931702
61
+ 600,0.0838,9.730426794137727e-06,1.227401614189148
62
+ 610,0.088,9.71687559561188e-06,0.9646913409233093
63
+ 620,0.0751,9.70300200490473e-06,0.9182428121566772
64
+ 630,0.0668,9.688806970257773e-06,0.9375871419906616
65
+ 640,0.071,9.674291461882774e-06,0.8192697763442993
66
+ 650,0.0719,9.659456471895446e-06,1.0246542692184448
67
+ 660,0.0704,9.64430301424765e-06,0.9603017568588257
68
+ 670,0.0841,9.628832124658085e-06,0.8320565819740295
69
+ 680,0.0721,9.613044860541508e-06,0.8893842101097107
70
+ 690,0.0666,9.596942300936445e-06,0.9222747087478638
71
+ 700,0.0686,9.580525546431458e-06,0.919222354888916
72
+ 710,0.078,9.563795719089913e-06,0.9898943901062012
73
+ 720,0.0659,9.546753962373281e-06,0.9966325163841248
74
+ 730,0.0674,9.529401441062998e-06,0.8561023473739624
75
+ 740,0.0559,9.511739341180843e-06,0.9110234379768372
76
+ 750,0.064,9.493768869907886e-06,0.884001612663269
77
+ 760,0.056,9.475491255501967e-06,0.877585232257843
78
+ 770,0.0579,9.456907747213748e-06,0.8948715925216675
79
+ 780,0.0598,9.438019615201338e-06,0.8685566782951355
80
+ 790,0.0565,9.418828150443469e-06,0.9344303607940674
81
+ 800,0.0619,9.399334664651263e-06,0.8488605618476868
82
+ 810,0.0629,9.379540490178582e-06,0.9291284084320068
83
+ 820,0.0585,9.359446979930954e-06,0.8899627327919006
84
+ 830,0.0596,9.33905550727312e-06,0.8248230814933777
85
+ 840,0.0542,9.318367465935143e-06,0.7917846441268921
86
+ 850,0.0647,9.297384269917169e-06,0.9088793992996216
87
+ 860,0.0562,9.276107353392774e-06,0.8969980478286743
88
+ 870,0.0572,9.254538170610938e-06,0.8161322474479675
89
+ 880,0.0511,9.232678195796655e-06,0.7354243397712708
90
+ 890,0.0498,9.210528923050164e-06,0.8445444107055664
91
+ 900,0.0594,9.188091866244833e-06,0.7331781387329102
92
+ 910,0.0475,9.165368558923695e-06,0.9718531966209412
93
+ 920,0.0534,9.142360554194617e-06,0.8079394102096558
94
+ 930,0.0487,9.119069424624165e-06,0.7201863527297974
95
+ 940,0.0466,9.095496762130099e-06,0.8865998387336731
96
+ 950,0.0576,9.071644177872594e-06,1.0389326810836792
97
+ 960,0.0444,9.047513302144096e-06,0.7415164709091187
98
+ 970,0.0467,9.023105784257906e-06,0.7516604661941528
99
+ 980,0.0496,8.998423292435455e-06,0.9304906725883484
100
+ 990,0.0527,8.973467513692265e-06,0.8555199503898621
101
+ 1000,0.0482,8.94824015372267e-06,0.7662856578826904
102
+ 1010,0.0457,8.922742936783207e-06,0.8362768292427063
103
+ 1020,0.0468,8.896977605574788e-06,0.8270362615585327
104
+ 1030,0.0476,8.870945921123576e-06,0.8668210506439209
105
+ 1040,0.0481,8.844649662660623e-06,0.7505970597267151
106
+ 1050,0.0421,8.818090627500266e-06,0.8698759078979492
107
+ 1060,0.0534,8.791270630917276e-06,0.7640476226806641
108
+ 1070,0.0486,8.764191506022794e-06,0.8142609000205994
109
+ 1080,0.0491,8.736855103639038e-06,0.6571221947669983
110
+ 1090,0.0361,8.709263292172796e-06,0.5830385684967041
111
+ 1100,0.043,8.68141795748773e-06,0.8238739371299744
112
+ 1110,0.0481,8.65332100277548e-06,0.8775807619094849
113
+ 1120,0.0465,8.624974348425575e-06,0.7803472876548767
114
+ 1130,0.0501,8.596379931894187e-06,0.8338977694511414
115
+ 1140,0.0493,8.567539707571704e-06,0.6809723973274231
116
+ 1150,0.0458,8.538455646649146e-06,0.779438316822052
117
+ 1160,0.044,8.509129736983446e-06,0.8775070309638977
118
+ 1170,0.042,8.479563982961572e-06,0.738409698009491
119
+ 1180,0.0433,8.449760405363538e-06,0.7935025095939636
120
+ 1190,0.0446,8.419721041224286e-06,0.9372588992118835
121
+ 1200,0.0456,8.389447943694451e-06,0.7028743028640747
122
+ 1210,0.0439,8.358943181900032e-06,0.7677087783813477
123
+ 1220,0.0374,8.328208840800981e-06,0.6402320861816406
124
+ 1230,0.0438,8.297247021048685e-06,0.9483605027198792
125
+ 1240,0.0392,8.266059838842396e-06,0.710944652557373
126
+ 1250,0.0454,8.23464942578459e-06,0.612249493598938
127
+ 1260,0.038,8.203017928735278e-06,0.6189194917678833
128
+ 1270,0.0442,8.171167509665261e-06,0.7650754451751709
129
+ 1280,0.0419,8.139100345508378e-06,0.71224445104599
130
+ 1290,0.0446,8.106818628012698e-06,0.9028244018554688
131
+ 1300,0.042,8.074324563590736e-06,0.746516227722168
132
+ 1310,0.0386,8.041620373168628e-06,0.6074661016464233
133
+ 1320,0.0427,8.00870829203435e-06,0.7320516705513
134
+ 1330,0.0369,7.975590569684926e-06,0.7389794588088989
135
+ 1340,0.0393,7.942269469672688e-06,0.6718634366989136
136
+ 1350,0.0342,7.908747269450558e-06,0.8955004811286926
137
+ 1360,0.0371,7.875026260216395e-06,0.7186753153800964
138
+ 1370,0.03,7.841108746756383e-06,0.7176233530044556
139
+ 1380,0.0349,7.806997047287515e-06,0.689918577671051
140
+ 1390,0.0319,7.772693493299139e-06,0.6836857199668884
141
+ 1400,0.0391,7.738200429393601e-06,0.8493905067443848
142
+ 1410,0.0317,7.703520213126e-06,0.8006901741027832
143
+ 1420,0.0389,7.66865521484305e-06,0.7174612879753113
144
+ 1430,0.0357,7.633607817521074e-06,0.6980575323104858
145
+ 1440,0.0362,7.59838041660312e-06,0.7103762626647949
146
+ 1450,0.0388,7.5629754198352475e-06,0.7503710985183716
147
+ 1460,0.0353,7.527395247101956e-06,0.7302154898643494
148
+ 1470,0.0359,7.491642330260789e-06,0.6966300010681152
149
+ 1480,0.0411,7.455719112976119e-06,0.6958418488502502
150
+ 1490,0.0327,7.419628050552131e-06,0.5858579874038696
151
+ 1500,0.0439,7.383371609764999e-06,0.7254495620727539
152
+ 1510,0.0343,7.3469522686942885e-06,0.7627992630004883
153
+ 1520,0.035,7.310372516553585e-06,0.7950798869132996
154
+ 1530,0.0362,7.273634853520357e-06,0.683579683303833
155
+ 1540,0.0413,7.2367417905650725e-06,0.6747159361839294
156
+ 1550,0.037,7.199695849279576e-06,0.7270603179931641
157
+ 1560,0.0334,7.1624995617047475e-06,0.6702049970626831
158
+ 1570,0.0321,7.12515547015743e-06,0.8103553056716919
159
+ 1580,0.032,7.0876661270566745e-06,0.7140473127365112
160
+ 1590,0.0345,7.050034094749286e-06,0.6398239135742188
161
+ 1600,0.0325,7.012261945334683e-06,0.7120059132575989
162
+ 1610,0.0291,6.974352260489104e-06,0.6111170053482056
163
+ 1620,0.0278,6.936307631289148e-06,0.6516941785812378
164
+ 1630,0.0291,6.898130658034685e-06,0.686859130859375
165
+ 1640,0.0324,6.859823950071127e-06,0.8052101135253906
166
+ 1650,0.036,6.8213901256110785e-06,0.7104664444923401
167
+ 1660,0.0252,6.782831811555386e-06,0.5812593698501587
168
+ 1670,0.0318,6.744151643313597e-06,0.7957876920700073
169
+ 1680,0.0337,6.705352264623829e-06,0.6455429196357727
170
+ 1690,0.0352,6.666436327372079e-06,0.6577746272087097
171
+ 1700,0.0312,6.627406491410961e-06,0.7138893008232117
172
+ 1710,0.0318,6.58826542437792e-06,0.6545776724815369
173
+ 1720,0.0349,6.549015801512895e-06,0.5184334516525269
174
+ 1730,0.0316,6.509660305475468e-06,0.6423726081848145
175
+ 1740,0.0279,6.47020162616152e-06,0.5985066890716553
176
+ 1750,0.0314,6.430642460519365e-06,0.7095875144004822
177
+ 1760,0.0351,6.390985512365426e-06,0.7356165051460266
178
+ 1770,0.0327,6.351233492199432e-06,0.6364956498146057
179
+ 1780,0.0295,6.311389117019155e-06,0.6260037422180176
180
+ 1790,0.0368,6.271455110134713e-06,0.770182192325592
181
+ 1800,0.0309,6.231434200982429e-06,0.6314713954925537
182
+ 1810,0.0303,6.191329124938284e-06,0.5799538493156433
183
+ 1820,0.0274,6.15114262313095e-06,0.6934332251548767
184
+ 1830,0.0276,6.110877442254445e-06,0.5117380023002625
185
+ 1840,0.0272,6.0705363343803945e-06,0.5944725275039673
186
+ 1850,0.0271,6.030122056769935e-06,0.7068486213684082
187
+ 1860,0.032,5.989637371685258e-06,0.6229940056800842
188
+ 1870,0.0319,5.949085046200808e-06,0.5604724884033203
189
+ 1880,0.0258,5.908467852014169e-06,0.6178894639015198
190
+ 1890,0.0264,5.867788565256607e-06,0.5318469405174255
191
+ 1900,0.0361,5.827049966303336e-06,0.5695937871932983
192
+ 1910,0.0313,5.786254839583478e-06,0.6509976983070374
193
+ 1920,0.0287,5.745405973389757e-06,0.8354060053825378
194
+ 1930,0.0297,5.704506159687914e-06,0.6091945171356201
195
+ 1940,0.0282,5.6635581939258855e-06,0.6996727585792542
196
+ 1950,0.0288,5.622564874842742e-06,0.6515761613845825
197
+ 1960,0.0232,5.581529004277384e-06,0.6472054719924927
198
+ 1970,0.0294,5.540453386977058e-06,0.6369526386260986
199
+ 1980,0.0357,5.4993408304056425e-06,0.6562472581863403
200
+ 1990,0.0267,5.458194144551769e-06,0.7148199677467346
201
+ 2000,0.0241,5.417016141736756e-06,0.5509341955184937
202
+ 2010,0.0228,5.3758096364223985e-06,0.503351092338562
203
+ 2020,0.0283,5.334577445018599e-06,0.6668328046798706
204
+ 2030,0.0303,5.293322385690868e-06,0.5887419581413269
205
+ 2040,0.0327,5.252047278167709e-06,0.5132143497467041
206
+ 2050,0.0267,5.210754943547893e-06,0.5204659700393677
207
+ 2060,0.0236,5.1694482041076434e-06,0.5391198396682739
208
+ 2070,0.0242,5.128129883107728e-06,0.36996734142303467
209
+ 2080,0.0219,5.086802804600505e-06,0.45624616742134094
210
+ 2090,0.0232,5.0454697932368925e-06,0.5221050977706909
211
+ 2100,0.0224,5.004133674073311e-06,0.6856546401977539
212
+ 2110,0.0223,4.962797272378597e-06,0.5804242491722107
213
+ 2120,0.0216,4.921463413440898e-06,0.5315669178962708
214
+ 2130,0.025,4.880134922374566e-06,0.48942509293556213
215
+ 2140,0.0251,4.838814623927067e-06,0.4987699091434479
216
+ 2150,0.0309,4.797505342285912e-06,0.5680892467498779
217
+ 2160,0.0246,4.756209900885628e-06,0.6083803176879883
218
+ 2170,0.0219,4.714931122214781e-06,0.4339863061904907
219
+ 2180,0.0277,4.673671827623058e-06,0.6728595495223999
220
+ 2190,0.0227,4.632434837128443e-06,0.5999028086662292
221
+ 2200,0.0235,4.591222969224453e-06,0.4993971586227417
222
+ 2210,0.0213,4.550039040687518e-06,0.4846261739730835
223
+ 2220,0.0214,4.508885866384446e-06,0.3943343758583069
224
+ 2230,0.0249,4.467766259080036e-06,0.6063972115516663
225
+ 2240,0.0208,4.426683029244825e-06,0.5530746579170227
226
+ 2250,0.0266,4.385638984863e-06,0.48010730743408203
227
+ 2260,0.0251,4.3446369312404745e-06,0.5987658500671387
228
+ 2270,0.0215,4.3036796708131475e-06,0.4759734869003296
229
+ 2280,0.0268,4.262770002955363e-06,0.6086810827255249
230
+ 2290,0.0241,4.221910723788578e-06,0.5050452351570129
231
+ 2300,0.0201,4.181104625990247e-06,0.5448557138442993
232
+ 2310,0.0289,4.140354498602952e-06,0.44929444789886475
233
+ 2320,0.0205,4.099663126843769e-06,0.4715081751346588
234
+ 2330,0.0306,4.059033291913902e-06,0.5584500432014465
235
+ 2340,0.0214,4.018467770808602e-06,0.6810476183891296
236
+ 2350,0.0266,3.9779693361273485e-06,0.5036186575889587
237
+ 2360,0.0218,3.937540755884357e-06,0.4518934488296509
238
+ 2370,0.0228,3.897184793319384e-06,0.4972905218601227
239
+ 2380,0.0199,3.856904206708864e-06,0.4766463041305542
240
+ 2390,0.0243,3.8167017491773845e-06,0.5434228181838989
241
+ 2400,0.0261,3.776580168509516e-06,0.7649377584457397
242
+ 2410,0.0209,3.7365422069620004e-06,0.5124170184135437
243
+ 2420,0.0216,3.6965906010763266e-06,0.482352077960968
244
+ 2430,0.0194,3.6567280814916863e-06,0.5959571599960327
245
+ 2440,0.0229,3.6169573727583408e-06,0.5238166451454163
246
+ 2450,0.0205,3.577281193151404e-06,0.46994876861572266
247
+ 2460,0.0249,3.5377022544850504e-06,0.713087797164917
248
+ 2470,0.0222,3.4982232619271583e-06,0.6384024620056152
249
+ 2480,0.0219,3.45884691381443e-06,0.42046940326690674
250
+ 2490,0.0167,3.419575901467952e-06,0.4439805746078491
251
+ 2500,0.0239,3.3804129090092542e-06,0.4470924437046051
252
+ 2510,0.0226,3.341360613176848e-06,0.48911812901496887
253
+ 2520,0.0206,3.302421683143279e-06,0.4699723720550537
254
+ 2530,0.0183,3.2635987803326898e-06,0.4276891052722931
255
+ 2540,0.0232,3.224894558238918e-06,0.5374788641929626
256
+ 2550,0.0266,3.18631166224413e-06,0.5477874279022217
257
+ 2560,0.0223,3.1478527294380167e-06,0.4970232844352722
258
+ 2570,0.0204,3.109520388437548e-06,0.4404042363166809
259
+ 2580,0.0224,3.0713172592073115e-06,0.5103297829627991
260
+ 2590,0.022,3.0332459528804457e-06,0.4886443316936493
261
+ 2600,0.0197,2.9953090715801637e-06,0.5699540972709656
262
+ 2610,0.0191,2.957509208241909e-06,0.42657411098480225
263
+ 2620,0.0202,2.919848946436129e-06,0.42806345224380493
264
+ 2630,0.0182,2.882330860191695e-06,0.3878917396068573
265
+ 2640,0.0165,2.8449575138199616e-06,0.4225691258907318
266
+ 2650,0.0221,2.8077314617395092e-06,0.4558577239513397
267
+ 2660,0.0286,2.7706552483015487e-06,0.4587768018245697
268
+ 2670,0.0225,2.733731407616018e-06,0.4155205190181732
269
+ 2680,0.0245,2.6969624633783807e-06,0.4576757848262787
270
+ 2690,0.0193,2.6603509286971342e-06,0.4659465551376343
271
+ 2700,0.0227,2.6238993059220397e-06,0.429494172334671
272
+ 2710,0.0218,2.5876100864730937e-06,0.39679718017578125
273
+ 2720,0.0194,2.5514857506702405e-06,0.47302669286727905
274
+ 2730,0.02,2.5155287675638475e-06,0.5222817659378052
275
+ 2740,0.0204,2.4797415947659457e-06,0.3765262961387634
276
+ 2750,0.0164,2.4441266782822585e-06,0.3531080186367035
277
+ 2760,0.0179,2.408686452345018e-06,0.44669318199157715
278
+ 2770,0.0186,2.3734233392465904e-06,0.347012996673584
279
+ 2780,0.0192,2.3383397491739147e-06,0.4109337329864502
280
+ 2790,0.0171,2.303438080043768e-06,0.4770718216896057
281
+ 2800,0.0191,2.2687207173388742e-06,0.37995004653930664
282
+ 2810,0.0204,2.234190033944858e-06,0.4003908038139343
283
+ 2820,0.0174,2.1998483899880595e-06,0.41007277369499207
284
+ 2830,0.0179,2.1656981326742267e-06,0.341975599527359
285
+ 2840,0.0196,2.1317415961280824e-06,0.447408527135849
286
+ 2850,0.0194,2.0979811012337942e-06,0.38129299879074097
287
+ 2860,0.0191,2.0644189554763416e-06,0.3637341856956482
288
+ 2870,0.0177,2.031057452783807e-06,0.4247613549232483
289
+ 2880,0.0166,1.9978988733705807e-06,0.3902631402015686
290
+ 2890,0.0173,1.9649454835815203e-06,0.4235377907752991
291
+ 2900,0.0181,1.932199535737045e-06,0.38576406240463257
292
+ 2910,0.0271,1.8996632679791915e-06,0.42286616563796997
293
+ 2920,0.0176,1.8673389041186418e-06,0.3406328856945038
294
+ 2930,0.0174,1.8352286534827273e-06,0.3646836280822754
295
+ 2940,0.0181,1.8033347107644261e-06,0.29196929931640625
296
+ 2950,0.017,1.7716592558723554e-06,0.37084370851516724
297
+ 2960,0.0221,1.7402044537817824e-06,0.3852077126502991
298
+ 2970,0.0212,1.7089724543866465e-06,0.35981303453445435
299
+ 2980,0.0162,1.677965392352619e-06,0.38413113355636597
300
+ 2990,0.0151,1.6471853869712023e-06,0.3001800775527954
301
+ 3000,0.02,1.6166345420148788e-06,0.3636736273765564
302
+ 3010,0.0175,1.5863149455933157e-06,0.43087366223335266
303
+ 3020,0.0226,1.556228670010656e-06,0.46356067061424255
304
+ 3030,0.0167,1.5263777716238693e-06,0.4455799460411072
305
+ 3040,0.0152,1.496764290702209e-06,0.32534974813461304
306
+ 3050,0.02,1.4673902512877585e-06,0.35252273082733154
307
+ 3060,0.0203,1.438257661057093e-06,0.38288936018943787
308
+ 3070,0.0214,1.4093685111840567e-06,0.3711300492286682
309
+ 3080,0.0177,1.3807247762036679e-06,0.35592737793922424
310
+ 3090,0.0184,1.3523284138771642e-06,0.33802157640457153
311
+ 3100,0.0187,1.3241813650581903e-06,0.3496468663215637
312
+ 3110,0.0215,1.2962855535601437e-06,0.2807375192642212
313
+ 3120,0.0167,1.2686428860246852e-06,0.2698224186897278
314
+ 3130,0.0194,1.241255251791421e-06,0.31049609184265137
315
+ 3140,0.0188,1.2141245227687732e-06,0.27130576968193054
316
+ 3150,0.0183,1.1872525533060269e-06,0.338786780834198
317
+ 3160,0.0163,1.1606411800666029e-06,0.32123616337776184
318
+ 3170,0.0196,1.134292221902511e-06,0.26763370633125305
319
+ 3180,0.0168,1.1082074797300413e-06,0.2890917658805847
320
+ 3190,0.0184,1.0823887364066737e-06,0.3752298057079315
321
+ 3200,0.0208,1.0568377566092163e-06,0.35502296686172485
322
+ 3210,0.0181,1.0315562867131984e-06,0.24394263327121735
323
+ 3220,0.0154,1.0065460546735045e-06,0.28185832500457764
324
+ 3230,0.0164,9.818087699062717e-07,0.25123700499534607
325
+ 3240,0.0139,9.573461231720549e-07,0.21169620752334595
326
+ 3250,0.0136,9.331597864602631e-07,0.2511177062988281
327
+ 3260,0.0233,9.09251412874882e-07,0.32039353251457214
328
+ 3270,0.0171,8.856226365214898e-07,0.25560271739959717
329
+ 3280,0.0143,8.622750723955597e-07,0.22130009531974792
330
+ 3290,0.0169,8.392103162720883e-07,0.2548290491104126
331
+ 3300,0.0152,8.164299445965168e-07,0.2960495948791504
332
+ 3310,0.016,7.939355143769905e-07,0.25807198882102966
333
+ 3320,0.0143,7.717285630779342e-07,0.23600825667381287
334
+ 3330,0.0175,7.498106085149698e-07,0.2826487421989441
335
+ 3340,0.015,7.281831487511775e-07,0.23631934821605682
336
+ 3350,0.021,7.06847661994704e-07,0.2791089415550232
337
+ 3360,0.0148,6.85805606497727e-07,0.24949561059474945
338
+ 3370,0.0246,6.650584204567889e-07,0.2484292834997177
339
+ 3380,0.0189,6.446075219144965e-07,0.2768440842628479
340
+ 3390,0.0196,6.244543086625987e-07,0.2492612600326538
341
+ 3400,0.0216,6.046001581464506e-07,0.24040444195270538
342
+ 3410,0.0168,5.850464273708689e-07,0.22290167212486267
343
+ 3420,0.0172,5.657944528073745e-07,0.2511428892612457
344
+ 3430,0.0156,5.468455503028574e-07,0.19910651445388794
345
+ 3440,0.0183,5.282010149896327e-07,0.25736910104751587
346
+ 3450,0.0145,5.098621211969224e-07,0.21390801668167114
347
+ 3460,0.0199,4.918301223637572e-07,0.21352550387382507
348
+ 3470,0.0144,4.74106250953304e-07,0.1626943051815033
349
+ 3480,0.0174,4.566917183686309e-07,0.20426668226718903
350
+ 3490,0.0161,4.395877148699074e-07,0.19366756081581116
351
+ 3500,0.0228,4.2279540949305197e-07,0.2561643123626709
352
+ 3510,0.0178,4.0631594996983104e-07,0.21617120504379272
353
+ 3520,0.0158,3.9015046264941346e-07,0.17024007439613342
354
+ 3530,0.0164,3.743000524213836e-07,0.1839084029197693
355
+ 3540,0.0214,3.587658026402263e-07,0.18981817364692688
356
+ 3550,0.017,3.435487750512778e-07,0.20637458562850952
357
+ 3560,0.0169,3.2865000971816107e-07,0.2143174260854721
358
+ 3570,0.019,3.140705249516957e-07,0.242491215467453
359
+ 3580,0.0187,2.998113172402989e-07,0.16586753726005554
360
+ 3590,0.0155,2.8587336118187647e-07,0.1478537917137146
361
+ 3600,0.0186,2.7225760941721136e-07,0.17264361679553986
362
+ 3610,0.0142,2.589649925648491e-07,0.14874860644340515
363
+ 3620,0.016,2.4599641915749486e-07,0.12994228303432465
364
+ 3630,0.017,2.3335277557991366e-07,0.17542079091072083
365
+ 3640,0.0181,2.210349260083494e-07,0.15290717780590057
366
+ 3650,0.0191,2.0904371235145826e-07,0.13746102154254913
367
+ 3660,0.0135,1.9737995419276456e-07,0.12954460084438324
368
+ 3670,0.0161,1.8604444873464468e-07,0.16208025813102722
369
+ 3680,0.0189,1.7503797074383988e-07,0.15591351687908173
370
+ 3690,0.0192,1.643612724985e-07,0.15741115808486938
371
+ 3700,0.0153,1.5401508373676766e-07,0.15939418971538544
372
+ 3710,0.0216,1.4400011160690175e-07,0.13837100565433502
373
+ 3720,0.0185,1.3431704061894313e-07,0.13348400592803955
374
+ 3730,0.0146,1.249665325979327e-07,0.15153057873249054
375
+ 3740,0.021,1.1594922663867137e-07,0.18070271611213684
376
+ 3750,0.0168,1.0726573906204463e-07,0.17874041199684143
377
+ 3760,0.0165,9.891666337289274e-08,0.12718811631202698
378
+ 3770,0.0214,9.090257021944881e-08,0.15776681900024414
379
+ 3780,0.0162,8.3224007354335e-08,0.13358241319656372
380
+ 3790,0.0149,7.588149959712243e-08,0.17998890578746796
381
+ 3800,0.0183,6.887554879846326e-08,0.12411261349916458
382
+ 3810,0.0143,6.220663380578862e-08,0.1386452466249466
383
+ 3820,0.0194,5.587521043057931e-08,0.10284613817930222
384
+ 3830,0.0196,4.988171141721232e-08,0.09959796071052551
385
+ 3840,0.0153,4.422654641338398e-08,0.11846427619457245
386
+ 3850,0.016,3.891010194211009e-08,0.12655889987945557
387
+ 3860,0.0189,3.3932741375308775e-08,0.11855187267065048
388
+ 3870,0.0174,2.9294804908962526e-08,0.1936274766921997
389
+ 3880,0.0167,2.499660953986849e-08,0.08700045198202133
390
+ 3890,0.0147,2.1038449043970232e-08,0.08493778109550476
391
+ 3900,0.0175,1.742059395627993e-08,0.11759188026189804
392
+ 3910,0.0161,1.4143291552387029e-08,0.12883485853672028
393
+ 3920,0.0155,1.1206765831557887e-08,0.10162201523780823
394
+ 3930,0.0163,8.611217501423575e-09,0.1062682569026947
395
+ 3940,0.0183,6.356823964266401e-09,0.15767623484134674
396
+ 3950,0.0193,4.4437393048885055e-09,0.11353452503681183
397
+ 3960,0.0149,2.8720942800858355e-09,0.11094997823238373
398
+ 3970,0.0178,1.6419963097080716e-09,0.09826391190290451
399
+ 3980,0.0152,7.535294693172823e-10,0.12383443862199783
400
+ 3990,0.0151,2.0675448444251734e-10,0.07030211389064789
401
+ 4000,0.0176,1.7087264264636915e-12,0.1488182246685028
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/model-00001-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:87f62f013113570ce2ce96d99e003444e1f7790bbdf74712f2d04fdad25f15c2
3
+ size 4990519232
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0406d79d34545813d621b61050c00d9ebcd45358b9465de7cdf4e6fb5f9c665e
3
+ size 1919980184
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/processor_config.json ADDED
@@ -0,0 +1,1154 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "processor_class": "Gr00tN1d7Processor",
3
+ "processor_kwargs": {
4
+ "modality_configs": {
5
+ "real_g1_relative_eef_relative_joints": {
6
+ "video": {
7
+ "delta_indices": [
8
+ -20,
9
+ 0
10
+ ],
11
+ "modality_keys": [
12
+ "ego_view"
13
+ ],
14
+ "sin_cos_embedding_keys": null,
15
+ "mean_std_embedding_keys": null,
16
+ "action_configs": null
17
+ },
18
+ "state": {
19
+ "delta_indices": [
20
+ 0
21
+ ],
22
+ "modality_keys": [
23
+ "left_wrist_eef_9d",
24
+ "right_wrist_eef_9d",
25
+ "left_hand",
26
+ "right_hand",
27
+ "left_arm",
28
+ "right_arm",
29
+ "waist"
30
+ ],
31
+ "sin_cos_embedding_keys": null,
32
+ "mean_std_embedding_keys": null,
33
+ "action_configs": null
34
+ },
35
+ "action": {
36
+ "delta_indices": [
37
+ 0,
38
+ 1,
39
+ 2,
40
+ 3,
41
+ 4,
42
+ 5,
43
+ 6,
44
+ 7,
45
+ 8,
46
+ 9,
47
+ 10,
48
+ 11,
49
+ 12,
50
+ 13,
51
+ 14,
52
+ 15,
53
+ 16,
54
+ 17,
55
+ 18,
56
+ 19,
57
+ 20,
58
+ 21,
59
+ 22,
60
+ 23,
61
+ 24,
62
+ 25,
63
+ 26,
64
+ 27,
65
+ 28,
66
+ 29,
67
+ 30,
68
+ 31,
69
+ 32,
70
+ 33,
71
+ 34,
72
+ 35,
73
+ 36,
74
+ 37,
75
+ 38,
76
+ 39
77
+ ],
78
+ "modality_keys": [
79
+ "left_wrist_eef_9d",
80
+ "right_wrist_eef_9d",
81
+ "left_hand",
82
+ "right_hand",
83
+ "left_arm",
84
+ "right_arm",
85
+ "waist",
86
+ "base_height_command",
87
+ "navigate_command"
88
+ ],
89
+ "sin_cos_embedding_keys": null,
90
+ "mean_std_embedding_keys": null,
91
+ "action_configs": [
92
+ {
93
+ "rep": "RELATIVE",
94
+ "type": "EEF",
95
+ "format": "XYZ_ROT6D",
96
+ "state_key": "left_wrist_eef_9d"
97
+ },
98
+ {
99
+ "rep": "RELATIVE",
100
+ "type": "EEF",
101
+ "format": "XYZ_ROT6D",
102
+ "state_key": "right_wrist_eef_9d"
103
+ },
104
+ {
105
+ "rep": "ABSOLUTE",
106
+ "type": "NON_EEF",
107
+ "format": "DEFAULT",
108
+ "state_key": "left_hand"
109
+ },
110
+ {
111
+ "rep": "ABSOLUTE",
112
+ "type": "NON_EEF",
113
+ "format": "DEFAULT",
114
+ "state_key": "right_hand"
115
+ },
116
+ {
117
+ "rep": "RELATIVE",
118
+ "type": "NON_EEF",
119
+ "format": "DEFAULT",
120
+ "state_key": "left_arm"
121
+ },
122
+ {
123
+ "rep": "RELATIVE",
124
+ "type": "NON_EEF",
125
+ "format": "DEFAULT",
126
+ "state_key": "right_arm"
127
+ },
128
+ {
129
+ "rep": "ABSOLUTE",
130
+ "type": "NON_EEF",
131
+ "format": "DEFAULT",
132
+ "state_key": "waist"
133
+ },
134
+ {
135
+ "rep": "ABSOLUTE",
136
+ "type": "NON_EEF",
137
+ "format": "DEFAULT",
138
+ "state_key": "base_height_command"
139
+ },
140
+ {
141
+ "rep": "ABSOLUTE",
142
+ "type": "NON_EEF",
143
+ "format": "DEFAULT",
144
+ "state_key": "navigate_command"
145
+ }
146
+ ]
147
+ },
148
+ "language": {
149
+ "delta_indices": [
150
+ 0
151
+ ],
152
+ "modality_keys": [
153
+ "annotation.human.task_description"
154
+ ],
155
+ "sin_cos_embedding_keys": null,
156
+ "mean_std_embedding_keys": null,
157
+ "action_configs": null
158
+ }
159
+ },
160
+ "real_r1_pro_sharpa_relative_eef_mecka": {
161
+ "video": {
162
+ "delta_indices": [
163
+ -30,
164
+ 0
165
+ ],
166
+ "modality_keys": [
167
+ "ego_view_cropratio_res320x240_freq30"
168
+ ],
169
+ "sin_cos_embedding_keys": null,
170
+ "mean_std_embedding_keys": null,
171
+ "action_configs": null
172
+ },
173
+ "state": {
174
+ "delta_indices": [
175
+ 0
176
+ ],
177
+ "modality_keys": [
178
+ "left_wrist_eef",
179
+ "right_wrist_eef",
180
+ "left_hand_joints",
181
+ "right_hand_joints"
182
+ ],
183
+ "sin_cos_embedding_keys": null,
184
+ "mean_std_embedding_keys": null,
185
+ "action_configs": null
186
+ },
187
+ "action": {
188
+ "delta_indices": [
189
+ 0,
190
+ 1,
191
+ 2,
192
+ 3,
193
+ 4,
194
+ 5,
195
+ 6,
196
+ 7,
197
+ 8,
198
+ 9,
199
+ 10,
200
+ 11,
201
+ 12,
202
+ 13,
203
+ 14,
204
+ 15,
205
+ 16,
206
+ 17,
207
+ 18,
208
+ 19,
209
+ 20,
210
+ 21,
211
+ 22,
212
+ 23,
213
+ 24,
214
+ 25,
215
+ 26,
216
+ 27,
217
+ 28,
218
+ 29,
219
+ 30,
220
+ 31,
221
+ 32,
222
+ 33,
223
+ 34,
224
+ 35,
225
+ 36,
226
+ 37,
227
+ 38,
228
+ 39
229
+ ],
230
+ "modality_keys": [
231
+ "left_wrist_eef",
232
+ "right_wrist_eef",
233
+ "left_hand_joints",
234
+ "right_hand_joints"
235
+ ],
236
+ "sin_cos_embedding_keys": null,
237
+ "mean_std_embedding_keys": null,
238
+ "action_configs": [
239
+ {
240
+ "rep": "RELATIVE",
241
+ "type": "EEF",
242
+ "format": "XYZ_ROT6D",
243
+ "state_key": "left_wrist_eef"
244
+ },
245
+ {
246
+ "rep": "RELATIVE",
247
+ "type": "EEF",
248
+ "format": "XYZ_ROT6D",
249
+ "state_key": "right_wrist_eef"
250
+ },
251
+ {
252
+ "rep": "ABSOLUTE",
253
+ "type": "NON_EEF",
254
+ "format": "DEFAULT",
255
+ "state_key": "left_hand_joints"
256
+ },
257
+ {
258
+ "rep": "ABSOLUTE",
259
+ "type": "NON_EEF",
260
+ "format": "DEFAULT",
261
+ "state_key": "right_hand_joints"
262
+ }
263
+ ]
264
+ },
265
+ "language": {
266
+ "delta_indices": [
267
+ 0
268
+ ],
269
+ "modality_keys": [
270
+ "annotation.human.coarse_action"
271
+ ],
272
+ "sin_cos_embedding_keys": null,
273
+ "mean_std_embedding_keys": null,
274
+ "action_configs": null
275
+ }
276
+ },
277
+ "real_r1_pro_sharpa_relative_eef_human": {
278
+ "video": {
279
+ "delta_indices": [
280
+ -20,
281
+ 0
282
+ ],
283
+ "modality_keys": [
284
+ "ego_view_res320x240_freq20",
285
+ "left_wrist_view_res320x240_freq20",
286
+ "right_wrist_view_res320x240_freq20"
287
+ ],
288
+ "sin_cos_embedding_keys": null,
289
+ "mean_std_embedding_keys": null,
290
+ "action_configs": null
291
+ },
292
+ "state": {
293
+ "delta_indices": [
294
+ 0
295
+ ],
296
+ "modality_keys": [
297
+ "left_wrist_eef",
298
+ "right_wrist_eef",
299
+ "left_hand_joints",
300
+ "right_hand_joints"
301
+ ],
302
+ "sin_cos_embedding_keys": null,
303
+ "mean_std_embedding_keys": null,
304
+ "action_configs": null
305
+ },
306
+ "action": {
307
+ "delta_indices": [
308
+ 0,
309
+ 1,
310
+ 2,
311
+ 3,
312
+ 4,
313
+ 5,
314
+ 6,
315
+ 7,
316
+ 8,
317
+ 9,
318
+ 10,
319
+ 11,
320
+ 12,
321
+ 13,
322
+ 14,
323
+ 15,
324
+ 16,
325
+ 17,
326
+ 18,
327
+ 19,
328
+ 20,
329
+ 21,
330
+ 22,
331
+ 23,
332
+ 24,
333
+ 25,
334
+ 26,
335
+ 27,
336
+ 28,
337
+ 29,
338
+ 30,
339
+ 31,
340
+ 32,
341
+ 33,
342
+ 34,
343
+ 35,
344
+ 36,
345
+ 37,
346
+ 38,
347
+ 39
348
+ ],
349
+ "modality_keys": [
350
+ "left_wrist_eef",
351
+ "right_wrist_eef",
352
+ "left_hand_joints",
353
+ "right_hand_joints"
354
+ ],
355
+ "sin_cos_embedding_keys": null,
356
+ "mean_std_embedding_keys": null,
357
+ "action_configs": [
358
+ {
359
+ "rep": "RELATIVE",
360
+ "type": "EEF",
361
+ "format": "XYZ_ROT6D",
362
+ "state_key": "left_wrist_eef"
363
+ },
364
+ {
365
+ "rep": "RELATIVE",
366
+ "type": "EEF",
367
+ "format": "XYZ_ROT6D",
368
+ "state_key": "right_wrist_eef"
369
+ },
370
+ {
371
+ "rep": "ABSOLUTE",
372
+ "type": "NON_EEF",
373
+ "format": "DEFAULT",
374
+ "state_key": "left_hand_joints"
375
+ },
376
+ {
377
+ "rep": "ABSOLUTE",
378
+ "type": "NON_EEF",
379
+ "format": "DEFAULT",
380
+ "state_key": "right_hand_joints"
381
+ }
382
+ ]
383
+ },
384
+ "language": {
385
+ "delta_indices": [
386
+ 0
387
+ ],
388
+ "modality_keys": [
389
+ "annotation.human.coarse_action"
390
+ ],
391
+ "sin_cos_embedding_keys": null,
392
+ "mean_std_embedding_keys": null,
393
+ "action_configs": null
394
+ }
395
+ },
396
+ "real_r1_pro_sharpa_relative_eef": {
397
+ "video": {
398
+ "delta_indices": [
399
+ -20,
400
+ 0
401
+ ],
402
+ "modality_keys": [
403
+ "ego_view_res320x240_freq20",
404
+ "left_wrist_view_res320x240_freq20",
405
+ "right_wrist_view_res320x240_freq20"
406
+ ],
407
+ "sin_cos_embedding_keys": null,
408
+ "mean_std_embedding_keys": null,
409
+ "action_configs": null
410
+ },
411
+ "state": {
412
+ "delta_indices": [
413
+ 0
414
+ ],
415
+ "modality_keys": [
416
+ "left_wrist_eef",
417
+ "right_wrist_eef",
418
+ "left_hand_joints",
419
+ "right_hand_joints"
420
+ ],
421
+ "sin_cos_embedding_keys": null,
422
+ "mean_std_embedding_keys": null,
423
+ "action_configs": null
424
+ },
425
+ "action": {
426
+ "delta_indices": [
427
+ 0,
428
+ 1,
429
+ 2,
430
+ 3,
431
+ 4,
432
+ 5,
433
+ 6,
434
+ 7,
435
+ 8,
436
+ 9,
437
+ 10,
438
+ 11,
439
+ 12,
440
+ 13,
441
+ 14,
442
+ 15,
443
+ 16,
444
+ 17,
445
+ 18,
446
+ 19,
447
+ 20,
448
+ 21,
449
+ 22,
450
+ 23,
451
+ 24,
452
+ 25,
453
+ 26,
454
+ 27,
455
+ 28,
456
+ 29,
457
+ 30,
458
+ 31,
459
+ 32,
460
+ 33,
461
+ 34,
462
+ 35,
463
+ 36,
464
+ 37,
465
+ 38,
466
+ 39
467
+ ],
468
+ "modality_keys": [
469
+ "left_wrist_eef",
470
+ "right_wrist_eef",
471
+ "left_hand_joints",
472
+ "right_hand_joints"
473
+ ],
474
+ "sin_cos_embedding_keys": null,
475
+ "mean_std_embedding_keys": null,
476
+ "action_configs": [
477
+ {
478
+ "rep": "RELATIVE",
479
+ "type": "EEF",
480
+ "format": "XYZ_ROT6D",
481
+ "state_key": "left_wrist_eef"
482
+ },
483
+ {
484
+ "rep": "RELATIVE",
485
+ "type": "EEF",
486
+ "format": "XYZ_ROT6D",
487
+ "state_key": "right_wrist_eef"
488
+ },
489
+ {
490
+ "rep": "ABSOLUTE",
491
+ "type": "NON_EEF",
492
+ "format": "DEFAULT",
493
+ "state_key": "left_hand_joints"
494
+ },
495
+ {
496
+ "rep": "ABSOLUTE",
497
+ "type": "NON_EEF",
498
+ "format": "DEFAULT",
499
+ "state_key": "right_hand_joints"
500
+ }
501
+ ]
502
+ },
503
+ "language": {
504
+ "delta_indices": [
505
+ 0
506
+ ],
507
+ "modality_keys": [
508
+ "annotation.human.coarse_action"
509
+ ],
510
+ "sin_cos_embedding_keys": null,
511
+ "mean_std_embedding_keys": null,
512
+ "action_configs": null
513
+ }
514
+ },
515
+ "xdof_relative_eef_relative_joint": {
516
+ "video": {
517
+ "delta_indices": [
518
+ -30,
519
+ 0
520
+ ],
521
+ "modality_keys": [
522
+ "top_camera-images-rgb_320_240",
523
+ "left_camera-images-rgb_320_240",
524
+ "right_camera-images-rgb_320_240"
525
+ ],
526
+ "sin_cos_embedding_keys": null,
527
+ "mean_std_embedding_keys": null,
528
+ "action_configs": null
529
+ },
530
+ "state": {
531
+ "delta_indices": [
532
+ 0
533
+ ],
534
+ "modality_keys": [
535
+ "left_wrist_eef",
536
+ "right_wrist_eef",
537
+ "left_gripper_pos",
538
+ "right_gripper_pos",
539
+ "left_joint_pos",
540
+ "right_joint_pos"
541
+ ],
542
+ "sin_cos_embedding_keys": null,
543
+ "mean_std_embedding_keys": null,
544
+ "action_configs": null
545
+ },
546
+ "action": {
547
+ "delta_indices": [
548
+ 0,
549
+ 1,
550
+ 2,
551
+ 3,
552
+ 4,
553
+ 5,
554
+ 6,
555
+ 7,
556
+ 8,
557
+ 9,
558
+ 10,
559
+ 11,
560
+ 12,
561
+ 13,
562
+ 14,
563
+ 15,
564
+ 16,
565
+ 17,
566
+ 18,
567
+ 19,
568
+ 20,
569
+ 21,
570
+ 22,
571
+ 23,
572
+ 24,
573
+ 25,
574
+ 26,
575
+ 27,
576
+ 28,
577
+ 29,
578
+ 30,
579
+ 31,
580
+ 32,
581
+ 33,
582
+ 34,
583
+ 35,
584
+ 36,
585
+ 37,
586
+ 38,
587
+ 39
588
+ ],
589
+ "modality_keys": [
590
+ "left_wrist_eef",
591
+ "right_wrist_eef",
592
+ "left_gripper_pos",
593
+ "right_gripper_pos",
594
+ "left_joint_pos",
595
+ "right_joint_pos"
596
+ ],
597
+ "sin_cos_embedding_keys": null,
598
+ "mean_std_embedding_keys": null,
599
+ "action_configs": [
600
+ {
601
+ "rep": "RELATIVE",
602
+ "type": "EEF",
603
+ "format": "XYZ_ROT6D",
604
+ "state_key": "left_wrist_eef"
605
+ },
606
+ {
607
+ "rep": "RELATIVE",
608
+ "type": "EEF",
609
+ "format": "XYZ_ROT6D",
610
+ "state_key": "right_wrist_eef"
611
+ },
612
+ {
613
+ "rep": "ABSOLUTE",
614
+ "type": "NON_EEF",
615
+ "format": "DEFAULT",
616
+ "state_key": "left_gripper_pos"
617
+ },
618
+ {
619
+ "rep": "ABSOLUTE",
620
+ "type": "NON_EEF",
621
+ "format": "DEFAULT",
622
+ "state_key": "right_gripper_pos"
623
+ },
624
+ {
625
+ "rep": "RELATIVE",
626
+ "type": "NON_EEF",
627
+ "format": "DEFAULT",
628
+ "state_key": "left_joint_pos"
629
+ },
630
+ {
631
+ "rep": "RELATIVE",
632
+ "type": "NON_EEF",
633
+ "format": "DEFAULT",
634
+ "state_key": "right_joint_pos"
635
+ }
636
+ ]
637
+ },
638
+ "language": {
639
+ "delta_indices": [
640
+ 0
641
+ ],
642
+ "modality_keys": [
643
+ "annotation.task"
644
+ ],
645
+ "sin_cos_embedding_keys": null,
646
+ "mean_std_embedding_keys": null,
647
+ "action_configs": null
648
+ }
649
+ },
650
+ "real_r1_pro_sharpa_relative_eef_maxinsights": {
651
+ "video": {
652
+ "delta_indices": [
653
+ -30,
654
+ 0
655
+ ],
656
+ "modality_keys": [
657
+ "ego_view_cropratio_res320x240_freq30"
658
+ ],
659
+ "sin_cos_embedding_keys": null,
660
+ "mean_std_embedding_keys": null,
661
+ "action_configs": null
662
+ },
663
+ "state": {
664
+ "delta_indices": [
665
+ 0
666
+ ],
667
+ "modality_keys": [
668
+ "left_wrist_eef",
669
+ "right_wrist_eef",
670
+ "left_hand_joints",
671
+ "right_hand_joints"
672
+ ],
673
+ "sin_cos_embedding_keys": null,
674
+ "mean_std_embedding_keys": null,
675
+ "action_configs": null
676
+ },
677
+ "action": {
678
+ "delta_indices": [
679
+ 0,
680
+ 1,
681
+ 2,
682
+ 3,
683
+ 4,
684
+ 5,
685
+ 6,
686
+ 7,
687
+ 8,
688
+ 9,
689
+ 10,
690
+ 11,
691
+ 12,
692
+ 13,
693
+ 14,
694
+ 15,
695
+ 16,
696
+ 17,
697
+ 18,
698
+ 19,
699
+ 20,
700
+ 21,
701
+ 22,
702
+ 23,
703
+ 24,
704
+ 25,
705
+ 26,
706
+ 27,
707
+ 28,
708
+ 29,
709
+ 30,
710
+ 31,
711
+ 32,
712
+ 33,
713
+ 34,
714
+ 35,
715
+ 36,
716
+ 37,
717
+ 38,
718
+ 39
719
+ ],
720
+ "modality_keys": [
721
+ "left_wrist_eef",
722
+ "right_wrist_eef",
723
+ "left_hand_joints",
724
+ "right_hand_joints"
725
+ ],
726
+ "sin_cos_embedding_keys": null,
727
+ "mean_std_embedding_keys": null,
728
+ "action_configs": [
729
+ {
730
+ "rep": "RELATIVE",
731
+ "type": "EEF",
732
+ "format": "XYZ_ROT6D",
733
+ "state_key": "left_wrist_eef"
734
+ },
735
+ {
736
+ "rep": "RELATIVE",
737
+ "type": "EEF",
738
+ "format": "XYZ_ROT6D",
739
+ "state_key": "right_wrist_eef"
740
+ },
741
+ {
742
+ "rep": "ABSOLUTE",
743
+ "type": "NON_EEF",
744
+ "format": "DEFAULT",
745
+ "state_key": "left_hand_joints"
746
+ },
747
+ {
748
+ "rep": "ABSOLUTE",
749
+ "type": "NON_EEF",
750
+ "format": "DEFAULT",
751
+ "state_key": "right_hand_joints"
752
+ }
753
+ ]
754
+ },
755
+ "language": {
756
+ "delta_indices": [
757
+ 0
758
+ ],
759
+ "modality_keys": [
760
+ "annotation.human.coarse_action"
761
+ ],
762
+ "sin_cos_embedding_keys": null,
763
+ "mean_std_embedding_keys": null,
764
+ "action_configs": null
765
+ }
766
+ },
767
+ "xdof_relative_eef_relative_joint_subtask": {
768
+ "video": {
769
+ "delta_indices": [
770
+ -30,
771
+ 0
772
+ ],
773
+ "modality_keys": [
774
+ "top_camera-images-rgb_320_240",
775
+ "left_camera-images-rgb_320_240",
776
+ "right_camera-images-rgb_320_240"
777
+ ],
778
+ "sin_cos_embedding_keys": null,
779
+ "mean_std_embedding_keys": null,
780
+ "action_configs": null
781
+ },
782
+ "state": {
783
+ "delta_indices": [
784
+ 0
785
+ ],
786
+ "modality_keys": [
787
+ "left_wrist_eef",
788
+ "right_wrist_eef",
789
+ "left_gripper_pos",
790
+ "right_gripper_pos",
791
+ "left_joint_pos",
792
+ "right_joint_pos"
793
+ ],
794
+ "sin_cos_embedding_keys": null,
795
+ "mean_std_embedding_keys": null,
796
+ "action_configs": null
797
+ },
798
+ "action": {
799
+ "delta_indices": [
800
+ 0,
801
+ 1,
802
+ 2,
803
+ 3,
804
+ 4,
805
+ 5,
806
+ 6,
807
+ 7,
808
+ 8,
809
+ 9,
810
+ 10,
811
+ 11,
812
+ 12,
813
+ 13,
814
+ 14,
815
+ 15,
816
+ 16,
817
+ 17,
818
+ 18,
819
+ 19,
820
+ 20,
821
+ 21,
822
+ 22,
823
+ 23,
824
+ 24,
825
+ 25,
826
+ 26,
827
+ 27,
828
+ 28,
829
+ 29,
830
+ 30,
831
+ 31,
832
+ 32,
833
+ 33,
834
+ 34,
835
+ 35,
836
+ 36,
837
+ 37,
838
+ 38,
839
+ 39
840
+ ],
841
+ "modality_keys": [
842
+ "left_wrist_eef",
843
+ "right_wrist_eef",
844
+ "left_gripper_pos",
845
+ "right_gripper_pos",
846
+ "left_joint_pos",
847
+ "right_joint_pos"
848
+ ],
849
+ "sin_cos_embedding_keys": null,
850
+ "mean_std_embedding_keys": null,
851
+ "action_configs": [
852
+ {
853
+ "rep": "RELATIVE",
854
+ "type": "EEF",
855
+ "format": "XYZ_ROT6D",
856
+ "state_key": "left_wrist_eef"
857
+ },
858
+ {
859
+ "rep": "RELATIVE",
860
+ "type": "EEF",
861
+ "format": "XYZ_ROT6D",
862
+ "state_key": "right_wrist_eef"
863
+ },
864
+ {
865
+ "rep": "ABSOLUTE",
866
+ "type": "NON_EEF",
867
+ "format": "DEFAULT",
868
+ "state_key": "left_gripper_pos"
869
+ },
870
+ {
871
+ "rep": "ABSOLUTE",
872
+ "type": "NON_EEF",
873
+ "format": "DEFAULT",
874
+ "state_key": "right_gripper_pos"
875
+ },
876
+ {
877
+ "rep": "RELATIVE",
878
+ "type": "NON_EEF",
879
+ "format": "DEFAULT",
880
+ "state_key": "left_joint_pos"
881
+ },
882
+ {
883
+ "rep": "RELATIVE",
884
+ "type": "NON_EEF",
885
+ "format": "DEFAULT",
886
+ "state_key": "right_joint_pos"
887
+ }
888
+ ]
889
+ },
890
+ "language": {
891
+ "delta_indices": [
892
+ 0
893
+ ],
894
+ "modality_keys": [
895
+ "annotation.sub_task"
896
+ ],
897
+ "sin_cos_embedding_keys": null,
898
+ "mean_std_embedding_keys": null,
899
+ "action_configs": null
900
+ }
901
+ },
902
+ "oxe_droid_relative_eef_relative_joint": {
903
+ "video": {
904
+ "delta_indices": [
905
+ -15,
906
+ 0
907
+ ],
908
+ "modality_keys": [
909
+ "exterior_image_1_left",
910
+ "wrist_image_left"
911
+ ],
912
+ "sin_cos_embedding_keys": null,
913
+ "mean_std_embedding_keys": null,
914
+ "action_configs": null
915
+ },
916
+ "state": {
917
+ "delta_indices": [
918
+ 0
919
+ ],
920
+ "modality_keys": [
921
+ "eef_9d",
922
+ "gripper_position",
923
+ "joint_position"
924
+ ],
925
+ "sin_cos_embedding_keys": null,
926
+ "mean_std_embedding_keys": null,
927
+ "action_configs": null
928
+ },
929
+ "action": {
930
+ "delta_indices": [
931
+ 0,
932
+ 1,
933
+ 2,
934
+ 3,
935
+ 4,
936
+ 5,
937
+ 6,
938
+ 7,
939
+ 8,
940
+ 9,
941
+ 10,
942
+ 11,
943
+ 12,
944
+ 13,
945
+ 14,
946
+ 15,
947
+ 16,
948
+ 17,
949
+ 18,
950
+ 19,
951
+ 20,
952
+ 21,
953
+ 22,
954
+ 23,
955
+ 24,
956
+ 25,
957
+ 26,
958
+ 27,
959
+ 28,
960
+ 29,
961
+ 30,
962
+ 31,
963
+ 32,
964
+ 33,
965
+ 34,
966
+ 35,
967
+ 36,
968
+ 37,
969
+ 38,
970
+ 39
971
+ ],
972
+ "modality_keys": [
973
+ "eef_9d",
974
+ "gripper_position",
975
+ "joint_position"
976
+ ],
977
+ "sin_cos_embedding_keys": null,
978
+ "mean_std_embedding_keys": null,
979
+ "action_configs": [
980
+ {
981
+ "rep": "RELATIVE",
982
+ "type": "EEF",
983
+ "format": "XYZ_ROT6D",
984
+ "state_key": "eef_9d"
985
+ },
986
+ {
987
+ "rep": "ABSOLUTE",
988
+ "type": "NON_EEF",
989
+ "format": "DEFAULT",
990
+ "state_key": "gripper_position"
991
+ },
992
+ {
993
+ "rep": "RELATIVE",
994
+ "type": "NON_EEF",
995
+ "format": "DEFAULT",
996
+ "state_key": "joint_position"
997
+ }
998
+ ]
999
+ },
1000
+ "language": {
1001
+ "delta_indices": [
1002
+ 0
1003
+ ],
1004
+ "modality_keys": [
1005
+ "annotation.language.language_instruction"
1006
+ ],
1007
+ "sin_cos_embedding_keys": null,
1008
+ "mean_std_embedding_keys": null,
1009
+ "action_configs": null
1010
+ }
1011
+ },
1012
+ "new_embodiment": {
1013
+ "video": {
1014
+ "delta_indices": [
1015
+ 0
1016
+ ],
1017
+ "modality_keys": [
1018
+ "head",
1019
+ "chest"
1020
+ ],
1021
+ "sin_cos_embedding_keys": null,
1022
+ "mean_std_embedding_keys": null,
1023
+ "action_configs": null
1024
+ },
1025
+ "state": {
1026
+ "delta_indices": [
1027
+ 0
1028
+ ],
1029
+ "modality_keys": [
1030
+ "left_leg",
1031
+ "right_leg",
1032
+ "waist",
1033
+ "left_arm",
1034
+ "right_arm",
1035
+ "head",
1036
+ "left_hand",
1037
+ "right_hand",
1038
+ "projected_gravity"
1039
+ ],
1040
+ "sin_cos_embedding_keys": null,
1041
+ "mean_std_embedding_keys": null,
1042
+ "action_configs": null
1043
+ },
1044
+ "action": {
1045
+ "delta_indices": [
1046
+ 0,
1047
+ 1,
1048
+ 2,
1049
+ 3,
1050
+ 4,
1051
+ 5,
1052
+ 6,
1053
+ 7,
1054
+ 8,
1055
+ 9,
1056
+ 10,
1057
+ 11,
1058
+ 12,
1059
+ 13,
1060
+ 14,
1061
+ 15,
1062
+ 16,
1063
+ 17,
1064
+ 18,
1065
+ 19,
1066
+ 20,
1067
+ 21,
1068
+ 22,
1069
+ 23,
1070
+ 24,
1071
+ 25,
1072
+ 26,
1073
+ 27,
1074
+ 28,
1075
+ 29,
1076
+ 30,
1077
+ 31
1078
+ ],
1079
+ "modality_keys": [
1080
+ "body_target",
1081
+ "left_hand_target",
1082
+ "right_hand_target"
1083
+ ],
1084
+ "sin_cos_embedding_keys": null,
1085
+ "mean_std_embedding_keys": null,
1086
+ "action_configs": [
1087
+ {
1088
+ "rep": "ABSOLUTE",
1089
+ "type": "NON_EEF",
1090
+ "format": "DEFAULT",
1091
+ "state_key": null
1092
+ },
1093
+ {
1094
+ "rep": "ABSOLUTE",
1095
+ "type": "NON_EEF",
1096
+ "format": "DEFAULT",
1097
+ "state_key": null
1098
+ },
1099
+ {
1100
+ "rep": "ABSOLUTE",
1101
+ "type": "NON_EEF",
1102
+ "format": "DEFAULT",
1103
+ "state_key": null
1104
+ }
1105
+ ]
1106
+ },
1107
+ "language": {
1108
+ "delta_indices": [
1109
+ 0
1110
+ ],
1111
+ "modality_keys": [
1112
+ "annotation.human.task_description"
1113
+ ],
1114
+ "sin_cos_embedding_keys": null,
1115
+ "mean_std_embedding_keys": null,
1116
+ "action_configs": null
1117
+ }
1118
+ }
1119
+ },
1120
+ "image_crop_size": [
1121
+ 230,
1122
+ 230
1123
+ ],
1124
+ "image_target_size": [
1125
+ 256,
1126
+ 256
1127
+ ],
1128
+ "use_albumentations": true,
1129
+ "random_rotation_angle": 0,
1130
+ "color_jitter_params": {
1131
+ "brightness": 0.3,
1132
+ "contrast": 0.4,
1133
+ "saturation": 0.5,
1134
+ "hue": 0.08
1135
+ },
1136
+ "shortest_image_edge": 256,
1137
+ "crop_fraction": 0.95,
1138
+ "letter_box_transform": false,
1139
+ "model_name": "nvidia/Cosmos-Reason2-2B",
1140
+ "model_type": "qwen",
1141
+ "formalize_language": true,
1142
+ "max_state_dim": 132,
1143
+ "max_action_dim": 132,
1144
+ "max_action_horizon": 40,
1145
+ "strict_action_padding_mask": false,
1146
+ "use_percentiles": true,
1147
+ "use_mean_std": false,
1148
+ "clip_outliers": true,
1149
+ "apply_sincos_state_encoding": false,
1150
+ "use_relative_action": true,
1151
+ "exclude_state": false,
1152
+ "state_dropout_prob": 0.0
1153
+ }
1154
+ }
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/protocol.json ADDED
@@ -0,0 +1,91 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "GR00T N1.7 (Gr00tN1d7), full fine-tune: all modules; variant 2026-11-15: every dropout 0, action horizon 32, 16 flow sampling steps, layer-wise lr (language 1e-6, vision 1e-5, vision-language merger and action head 1e-4)",
3
+ "task": "A18 pour water (H2), 2 cameras",
4
+ "task_dir": "a18-pour-water/b512-2cam-h32-cropped-v2",
5
+ "robot": "H2 (Sharpa hands)",
6
+ "training_data": "humanoid-benchmark/a18-pour-water-expert-cropped-v2 @ 44cdd5715f4681038bd0bdf377e7705585a831c3 (100 episodes, 66,850 frames; v1 @8e956a10 with the waits trimmed)",
7
+ "episodes": 100,
8
+ "epoch_completed": null,
9
+ "global_step": 2000,
10
+ "augmentation": "Humanoid Benchmark online augmentation spec v2 (2026-11-09): 15 visual effects, one effect per training sample at 1.5x; proprioception aligned with the GR00T A20 recipe (2026-11-12): per-sample joint bias +-2 deg + Gaussian noise 0.25 deg (hb_aug_v2.perturb_joints), no proprio delay, no history drop; model-native image augmentation off",
11
+ "epochs": null,
12
+ "total_steps": 4000,
13
+ "save_steps": 2000,
14
+ "global_batch": 512,
15
+ "per_gpu_batch": 32,
16
+ "num_gpus": 16,
17
+ "lr": 0.0001,
18
+ "lr_schedule": "cosine, warmup_ratio 0.05",
19
+ "weight_decay": 1e-05,
20
+ "proprio_augmentation": "per-sample joint bias +-2 deg + noise 0.25 deg (hb_aug_v2.perturb_joints); no proprio delay / history drop (the GR00T A20 recipe)",
21
+ "config": "experiment_cfg/conf.yaml",
22
+ "normalization": "statistics.json",
23
+ "inference": "NVIDIA Isaac-GR00T N1.7 policy loader with embodiment_tag new_embodiment and experiment_cfg/conf.yaml; VLM nvidia/Cosmos-Reason2-2B from its own repository.",
24
+ "cameras": [
25
+ "observation.images.head",
26
+ "observation.images.chest"
27
+ ],
28
+ "state_layout": "state78 = observation.state (75) + observation.projected_gravity (3); no joint velocity",
29
+ "action_layout": "action75 = action.body_target (31) + action.hands (44); 32 rows per chunk (horizon 32)",
30
+ "initialization": "GR00T N1.7 base (groot-n17-base); VLM nvidia/Cosmos-Reason2-2B",
31
+ "trainable": "all modules (tune_llm, tune_visual, projector, diffusion action head, VLLN)",
32
+ "recipe_reference": "groot/a18-pour-water/b256/step-004000 recipe (head + chest, 2 x 8 x 32 = 512, no accumulation, 4000 steps) with: state_dropout_prob 0, action-head attn_dropout 0, DiT dropout 0, action horizon 32, num_inference_timesteps 16, layer-wise lr (user 2026-11-15)",
33
+ "model_overrides": {
34
+ "attn_dropout": 0.0,
35
+ "diffusion_model_cfg.dropout": 0.0,
36
+ "state_dropout_prob": 0.0,
37
+ "num_inference_timesteps": 16
38
+ },
39
+ "lr_groups": {
40
+ "language": 1e-06,
41
+ "vision": 1e-05,
42
+ "projector": 0.0001,
43
+ "rest": 0.0001
44
+ },
45
+ "launcher": "code/hb_groot/hb_launch_v.py + groot_pair2x32_h2v.sh, modality h2_body31_hand44_h32.py",
46
+ "published_steps": [
47
+ 2000,
48
+ 4000
49
+ ],
50
+ "final_loss_mean_last50": 0.0172,
51
+ "loss_curve": "loss_curve.csv",
52
+ "training_records_included": true,
53
+ "resume_state_included": false,
54
+ "files": {
55
+ "code/hb_groot/cache_cosmos_reason2.py": 2542,
56
+ "code/hb_groot/check_reason2_offline.py": 1283,
57
+ "code/hb_groot/g1_feedback76_h16.py": 1474,
58
+ "code/hb_groot/groot_chain.py": 6325,
59
+ "code/hb_groot/h2_body31_hand44_h16.py": 1395,
60
+ "code/hb_groot/h2_body31_hand44_h32.py": 1463,
61
+ "code/hb_groot/h2_body31_hand44_planner7_h16.py": 1531,
62
+ "code/hb_groot/h2w3_body31_hand44_h16.py": 1563,
63
+ "code/hb_groot/hb_aug_v2.py": 7475,
64
+ "code/hb_groot/hb_inloop_eval.py": 16620,
65
+ "code/hb_groot/hb_launch.py": 4340,
66
+ "code/hb_groot/hb_launch_v.py": 4184,
67
+ "code/hb_groot/hb_offline_fix.py": 1138,
68
+ "code/hb_groot/prepare_groot_data.py": 13912,
69
+ "code/hb_groot/verify_groot_cpu.py": 2764,
70
+ "config.json": 2219,
71
+ "dataset_meta/hb-dataset-receipt.json": 789,
72
+ "dataset_meta/modality.json": 1549,
73
+ "embodiment_id.json": 2226,
74
+ "experiment_cfg/conf.yaml": 6017,
75
+ "experiment_cfg/config.yaml": 5985,
76
+ "experiment_cfg/dataset_statistics.json": 28317,
77
+ "experiment_cfg/final_model_config.json": 1584,
78
+ "experiment_cfg/final_processor_config.json": 2548463,
79
+ "loss_curve.csv": 21108,
80
+ "model-00001-of-00002.safetensors": 4990519232,
81
+ "model-00002-of-00002.safetensors": 1919980184,
82
+ "model.safetensors.index.json": 104985,
83
+ "processor_config.json": 27216,
84
+ "statistics.json": 4321410,
85
+ "train.log": 398701,
86
+ "trainer_state.json": 28303,
87
+ "training_args.bin": 8017,
88
+ "training_plan.json": 1004,
89
+ "wandb_config.json": 58
90
+ }
91
+ }
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/statistics.json ADDED
The diff for this file is too large to render. See raw diff
 
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/train.log ADDED
The diff for this file is too large to render. See raw diff
 
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/trainer_state.json ADDED
@@ -0,0 +1,1234 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_global_step": null,
3
+ "best_metric": null,
4
+ "best_model_checkpoint": null,
5
+ "epoch": 0.5,
6
+ "eval_steps": 500,
7
+ "global_step": 2000,
8
+ "is_hyper_param_search": false,
9
+ "is_local_process_zero": true,
10
+ "is_world_process_zero": true,
11
+ "log_history": [
12
+ {
13
+ "grad_norm": 0.24912892282009125,
14
+ "learning_rate": 4.5000000000000003e-07,
15
+ "loss": 1.3905,
16
+ "step": 10
17
+ },
18
+ {
19
+ "grad_norm": 0.394233375787735,
20
+ "learning_rate": 9.500000000000001e-07,
21
+ "loss": 1.3802,
22
+ "step": 20
23
+ },
24
+ {
25
+ "grad_norm": 1.497800588607788,
26
+ "learning_rate": 1.45e-06,
27
+ "loss": 1.3083,
28
+ "step": 30
29
+ },
30
+ {
31
+ "grad_norm": 1.1424145698547363,
32
+ "learning_rate": 1.9500000000000004e-06,
33
+ "loss": 1.2031,
34
+ "step": 40
35
+ },
36
+ {
37
+ "grad_norm": 1.2752643823623657,
38
+ "learning_rate": 2.4500000000000003e-06,
39
+ "loss": 1.1098,
40
+ "step": 50
41
+ },
42
+ {
43
+ "grad_norm": 1.0560007095336914,
44
+ "learning_rate": 2.95e-06,
45
+ "loss": 1.079,
46
+ "step": 60
47
+ },
48
+ {
49
+ "grad_norm": 0.720940351486206,
50
+ "learning_rate": 3.45e-06,
51
+ "loss": 1.0589,
52
+ "step": 70
53
+ },
54
+ {
55
+ "grad_norm": 0.4508174955844879,
56
+ "learning_rate": 3.95e-06,
57
+ "loss": 1.043,
58
+ "step": 80
59
+ },
60
+ {
61
+ "grad_norm": 0.4460732042789459,
62
+ "learning_rate": 4.450000000000001e-06,
63
+ "loss": 1.0379,
64
+ "step": 90
65
+ },
66
+ {
67
+ "grad_norm": 0.2577953338623047,
68
+ "learning_rate": 4.95e-06,
69
+ "loss": 1.0347,
70
+ "step": 100
71
+ },
72
+ {
73
+ "grad_norm": 0.5169486403465271,
74
+ "learning_rate": 5.450000000000001e-06,
75
+ "loss": 1.032,
76
+ "step": 110
77
+ },
78
+ {
79
+ "grad_norm": 0.44850045442581177,
80
+ "learning_rate": 5.950000000000001e-06,
81
+ "loss": 1.0248,
82
+ "step": 120
83
+ },
84
+ {
85
+ "grad_norm": 0.5352460145950317,
86
+ "learning_rate": 6.450000000000001e-06,
87
+ "loss": 1.0148,
88
+ "step": 130
89
+ },
90
+ {
91
+ "grad_norm": 0.738865315914154,
92
+ "learning_rate": 6.95e-06,
93
+ "loss": 0.9866,
94
+ "step": 140
95
+ },
96
+ {
97
+ "grad_norm": 0.6277132034301758,
98
+ "learning_rate": 7.450000000000001e-06,
99
+ "loss": 0.9231,
100
+ "step": 150
101
+ },
102
+ {
103
+ "grad_norm": 0.5653408169746399,
104
+ "learning_rate": 7.950000000000002e-06,
105
+ "loss": 0.8618,
106
+ "step": 160
107
+ },
108
+ {
109
+ "grad_norm": 0.5750411152839661,
110
+ "learning_rate": 8.45e-06,
111
+ "loss": 0.8111,
112
+ "step": 170
113
+ },
114
+ {
115
+ "grad_norm": 0.7955657839775085,
116
+ "learning_rate": 8.95e-06,
117
+ "loss": 0.7641,
118
+ "step": 180
119
+ },
120
+ {
121
+ "grad_norm": 0.8227806091308594,
122
+ "learning_rate": 9.450000000000001e-06,
123
+ "loss": 0.7216,
124
+ "step": 190
125
+ },
126
+ {
127
+ "grad_norm": 0.6388406753540039,
128
+ "learning_rate": 9.950000000000001e-06,
129
+ "loss": 0.6926,
130
+ "step": 200
131
+ },
132
+ {
133
+ "grad_norm": 0.6702894568443298,
134
+ "learning_rate": 9.999861593790126e-06,
135
+ "loss": 0.6776,
136
+ "step": 210
137
+ },
138
+ {
139
+ "grad_norm": 0.6838157176971436,
140
+ "learning_rate": 9.999383162408303e-06,
141
+ "loss": 0.6523,
142
+ "step": 220
143
+ },
144
+ {
145
+ "grad_norm": 0.6812769770622253,
146
+ "learning_rate": 9.998563029828258e-06,
147
+ "loss": 0.6323,
148
+ "step": 230
149
+ },
150
+ {
151
+ "grad_norm": 0.6650444865226746,
152
+ "learning_rate": 9.997401252104962e-06,
153
+ "loss": 0.6057,
154
+ "step": 240
155
+ },
156
+ {
157
+ "grad_norm": 1.195233702659607,
158
+ "learning_rate": 9.995897908644378e-06,
159
+ "loss": 0.5821,
160
+ "step": 250
161
+ },
162
+ {
163
+ "grad_norm": 0.8008884191513062,
164
+ "learning_rate": 9.994053102198035e-06,
165
+ "loss": 0.5615,
166
+ "step": 260
167
+ },
168
+ {
169
+ "grad_norm": 0.744047999382019,
170
+ "learning_rate": 9.991866958856003e-06,
171
+ "loss": 0.5517,
172
+ "step": 270
173
+ },
174
+ {
175
+ "grad_norm": 0.7080017328262329,
176
+ "learning_rate": 9.989339628038277e-06,
177
+ "loss": 0.5187,
178
+ "step": 280
179
+ },
180
+ {
181
+ "grad_norm": 0.7555558085441589,
182
+ "learning_rate": 9.98647128248456e-06,
183
+ "loss": 0.5128,
184
+ "step": 290
185
+ },
186
+ {
187
+ "grad_norm": 0.7810933589935303,
188
+ "learning_rate": 9.983262118242461e-06,
189
+ "loss": 0.4822,
190
+ "step": 300
191
+ },
192
+ {
193
+ "grad_norm": 0.8074049949645996,
194
+ "learning_rate": 9.979712354654092e-06,
195
+ "loss": 0.4704,
196
+ "step": 310
197
+ },
198
+ {
199
+ "grad_norm": 0.8137549757957458,
200
+ "learning_rate": 9.97582223434108e-06,
201
+ "loss": 0.4418,
202
+ "step": 320
203
+ },
204
+ {
205
+ "grad_norm": 0.8988544940948486,
206
+ "learning_rate": 9.971592023187979e-06,
207
+ "loss": 0.4221,
208
+ "step": 330
209
+ },
210
+ {
211
+ "grad_norm": 0.8715605139732361,
212
+ "learning_rate": 9.967022010324105e-06,
213
+ "loss": 0.3992,
214
+ "step": 340
215
+ },
216
+ {
217
+ "grad_norm": 0.9414624571800232,
218
+ "learning_rate": 9.962112508103766e-06,
219
+ "loss": 0.3715,
220
+ "step": 350
221
+ },
222
+ {
223
+ "grad_norm": 1.1080490350723267,
224
+ "learning_rate": 9.956863852084914e-06,
225
+ "loss": 0.3612,
226
+ "step": 360
227
+ },
228
+ {
229
+ "grad_norm": 0.9747459292411804,
230
+ "learning_rate": 9.95127640100622e-06,
231
+ "loss": 0.3463,
232
+ "step": 370
233
+ },
234
+ {
235
+ "grad_norm": 0.8970965147018433,
236
+ "learning_rate": 9.945350536762544e-06,
237
+ "loss": 0.3182,
238
+ "step": 380
239
+ },
240
+ {
241
+ "grad_norm": 1.0791109800338745,
242
+ "learning_rate": 9.93908666437883e-06,
243
+ "loss": 0.304,
244
+ "step": 390
245
+ },
246
+ {
247
+ "grad_norm": 1.1977298259735107,
248
+ "learning_rate": 9.932485211982437e-06,
249
+ "loss": 0.2832,
250
+ "step": 400
251
+ },
252
+ {
253
+ "grad_norm": 1.120720624923706,
254
+ "learning_rate": 9.92554663077387e-06,
255
+ "loss": 0.272,
256
+ "step": 410
257
+ },
258
+ {
259
+ "grad_norm": 1.2665119171142578,
260
+ "learning_rate": 9.918271394995936e-06,
261
+ "loss": 0.2509,
262
+ "step": 420
263
+ },
264
+ {
265
+ "grad_norm": 1.0502629280090332,
266
+ "learning_rate": 9.910660001901335e-06,
267
+ "loss": 0.2406,
268
+ "step": 430
269
+ },
270
+ {
271
+ "grad_norm": 1.1278620958328247,
272
+ "learning_rate": 9.902712971718676e-06,
273
+ "loss": 0.2187,
274
+ "step": 440
275
+ },
276
+ {
277
+ "grad_norm": 0.9573274254798889,
278
+ "learning_rate": 9.894430847616916e-06,
279
+ "loss": 0.2001,
280
+ "step": 450
281
+ },
282
+ {
283
+ "grad_norm": 0.9732072949409485,
284
+ "learning_rate": 9.885814195668232e-06,
285
+ "loss": 0.1847,
286
+ "step": 460
287
+ },
288
+ {
289
+ "grad_norm": 1.0263832807540894,
290
+ "learning_rate": 9.876863604809345e-06,
291
+ "loss": 0.1746,
292
+ "step": 470
293
+ },
294
+ {
295
+ "grad_norm": 1.1131765842437744,
296
+ "learning_rate": 9.867579686801245e-06,
297
+ "loss": 0.1565,
298
+ "step": 480
299
+ },
300
+ {
301
+ "grad_norm": 1.0001071691513062,
302
+ "learning_rate": 9.8579630761874e-06,
303
+ "loss": 0.143,
304
+ "step": 490
305
+ },
306
+ {
307
+ "grad_norm": 1.2044321298599243,
308
+ "learning_rate": 9.848014430250368e-06,
309
+ "loss": 0.1289,
310
+ "step": 500
311
+ },
312
+ {
313
+ "grad_norm": 1.260959267616272,
314
+ "learning_rate": 9.837734428966885e-06,
315
+ "loss": 0.1252,
316
+ "step": 510
317
+ },
318
+ {
319
+ "grad_norm": 1.1621663570404053,
320
+ "learning_rate": 9.827123774961383e-06,
321
+ "loss": 0.1066,
322
+ "step": 520
323
+ },
324
+ {
325
+ "grad_norm": 1.0683119297027588,
326
+ "learning_rate": 9.81618319345797e-06,
327
+ "loss": 0.1068,
328
+ "step": 530
329
+ },
330
+ {
331
+ "grad_norm": 1.4515645503997803,
332
+ "learning_rate": 9.804913432230856e-06,
333
+ "loss": 0.0993,
334
+ "step": 540
335
+ },
336
+ {
337
+ "grad_norm": 0.8918207287788391,
338
+ "learning_rate": 9.793315261553252e-06,
339
+ "loss": 0.0915,
340
+ "step": 550
341
+ },
342
+ {
343
+ "grad_norm": 0.9777098298072815,
344
+ "learning_rate": 9.781389474144718e-06,
345
+ "loss": 0.0821,
346
+ "step": 560
347
+ },
348
+ {
349
+ "grad_norm": 0.9528394937515259,
350
+ "learning_rate": 9.769136885116981e-06,
351
+ "loss": 0.0792,
352
+ "step": 570
353
+ },
354
+ {
355
+ "grad_norm": 0.9681083559989929,
356
+ "learning_rate": 9.756558331918227e-06,
357
+ "loss": 0.0809,
358
+ "step": 580
359
+ },
360
+ {
361
+ "grad_norm": 0.7851548790931702,
362
+ "learning_rate": 9.743654674275855e-06,
363
+ "loss": 0.0822,
364
+ "step": 590
365
+ },
366
+ {
367
+ "grad_norm": 1.227401614189148,
368
+ "learning_rate": 9.730426794137727e-06,
369
+ "loss": 0.0838,
370
+ "step": 600
371
+ },
372
+ {
373
+ "grad_norm": 0.9646913409233093,
374
+ "learning_rate": 9.71687559561188e-06,
375
+ "loss": 0.088,
376
+ "step": 610
377
+ },
378
+ {
379
+ "grad_norm": 0.9182428121566772,
380
+ "learning_rate": 9.70300200490473e-06,
381
+ "loss": 0.0751,
382
+ "step": 620
383
+ },
384
+ {
385
+ "grad_norm": 0.9375871419906616,
386
+ "learning_rate": 9.688806970257773e-06,
387
+ "loss": 0.0668,
388
+ "step": 630
389
+ },
390
+ {
391
+ "grad_norm": 0.8192697763442993,
392
+ "learning_rate": 9.674291461882774e-06,
393
+ "loss": 0.071,
394
+ "step": 640
395
+ },
396
+ {
397
+ "grad_norm": 1.0246542692184448,
398
+ "learning_rate": 9.659456471895446e-06,
399
+ "loss": 0.0719,
400
+ "step": 650
401
+ },
402
+ {
403
+ "grad_norm": 0.9603017568588257,
404
+ "learning_rate": 9.64430301424765e-06,
405
+ "loss": 0.0704,
406
+ "step": 660
407
+ },
408
+ {
409
+ "grad_norm": 0.8320565819740295,
410
+ "learning_rate": 9.628832124658085e-06,
411
+ "loss": 0.0841,
412
+ "step": 670
413
+ },
414
+ {
415
+ "grad_norm": 0.8893842101097107,
416
+ "learning_rate": 9.613044860541508e-06,
417
+ "loss": 0.0721,
418
+ "step": 680
419
+ },
420
+ {
421
+ "grad_norm": 0.9222747087478638,
422
+ "learning_rate": 9.596942300936445e-06,
423
+ "loss": 0.0666,
424
+ "step": 690
425
+ },
426
+ {
427
+ "grad_norm": 0.919222354888916,
428
+ "learning_rate": 9.580525546431458e-06,
429
+ "loss": 0.0686,
430
+ "step": 700
431
+ },
432
+ {
433
+ "grad_norm": 0.9898943901062012,
434
+ "learning_rate": 9.563795719089913e-06,
435
+ "loss": 0.078,
436
+ "step": 710
437
+ },
438
+ {
439
+ "grad_norm": 0.9966325163841248,
440
+ "learning_rate": 9.546753962373281e-06,
441
+ "loss": 0.0659,
442
+ "step": 720
443
+ },
444
+ {
445
+ "grad_norm": 0.8561023473739624,
446
+ "learning_rate": 9.529401441062998e-06,
447
+ "loss": 0.0674,
448
+ "step": 730
449
+ },
450
+ {
451
+ "grad_norm": 0.9110234379768372,
452
+ "learning_rate": 9.511739341180843e-06,
453
+ "loss": 0.0559,
454
+ "step": 740
455
+ },
456
+ {
457
+ "grad_norm": 0.884001612663269,
458
+ "learning_rate": 9.493768869907886e-06,
459
+ "loss": 0.064,
460
+ "step": 750
461
+ },
462
+ {
463
+ "grad_norm": 0.877585232257843,
464
+ "learning_rate": 9.475491255501967e-06,
465
+ "loss": 0.056,
466
+ "step": 760
467
+ },
468
+ {
469
+ "grad_norm": 0.8948715925216675,
470
+ "learning_rate": 9.456907747213748e-06,
471
+ "loss": 0.0579,
472
+ "step": 770
473
+ },
474
+ {
475
+ "grad_norm": 0.8685566782951355,
476
+ "learning_rate": 9.438019615201338e-06,
477
+ "loss": 0.0598,
478
+ "step": 780
479
+ },
480
+ {
481
+ "grad_norm": 0.9344303607940674,
482
+ "learning_rate": 9.418828150443469e-06,
483
+ "loss": 0.0565,
484
+ "step": 790
485
+ },
486
+ {
487
+ "grad_norm": 0.8488605618476868,
488
+ "learning_rate": 9.399334664651263e-06,
489
+ "loss": 0.0619,
490
+ "step": 800
491
+ },
492
+ {
493
+ "grad_norm": 0.9291284084320068,
494
+ "learning_rate": 9.379540490178582e-06,
495
+ "loss": 0.0629,
496
+ "step": 810
497
+ },
498
+ {
499
+ "grad_norm": 0.8899627327919006,
500
+ "learning_rate": 9.359446979930954e-06,
501
+ "loss": 0.0585,
502
+ "step": 820
503
+ },
504
+ {
505
+ "grad_norm": 0.8248230814933777,
506
+ "learning_rate": 9.33905550727312e-06,
507
+ "loss": 0.0596,
508
+ "step": 830
509
+ },
510
+ {
511
+ "grad_norm": 0.7917846441268921,
512
+ "learning_rate": 9.318367465935143e-06,
513
+ "loss": 0.0542,
514
+ "step": 840
515
+ },
516
+ {
517
+ "grad_norm": 0.9088793992996216,
518
+ "learning_rate": 9.297384269917169e-06,
519
+ "loss": 0.0647,
520
+ "step": 850
521
+ },
522
+ {
523
+ "grad_norm": 0.8969980478286743,
524
+ "learning_rate": 9.276107353392774e-06,
525
+ "loss": 0.0562,
526
+ "step": 860
527
+ },
528
+ {
529
+ "grad_norm": 0.8161322474479675,
530
+ "learning_rate": 9.254538170610938e-06,
531
+ "loss": 0.0572,
532
+ "step": 870
533
+ },
534
+ {
535
+ "grad_norm": 0.7354243397712708,
536
+ "learning_rate": 9.232678195796655e-06,
537
+ "loss": 0.0511,
538
+ "step": 880
539
+ },
540
+ {
541
+ "grad_norm": 0.8445444107055664,
542
+ "learning_rate": 9.210528923050164e-06,
543
+ "loss": 0.0498,
544
+ "step": 890
545
+ },
546
+ {
547
+ "grad_norm": 0.7331781387329102,
548
+ "learning_rate": 9.188091866244833e-06,
549
+ "loss": 0.0594,
550
+ "step": 900
551
+ },
552
+ {
553
+ "grad_norm": 0.9718531966209412,
554
+ "learning_rate": 9.165368558923695e-06,
555
+ "loss": 0.0475,
556
+ "step": 910
557
+ },
558
+ {
559
+ "grad_norm": 0.8079394102096558,
560
+ "learning_rate": 9.142360554194617e-06,
561
+ "loss": 0.0534,
562
+ "step": 920
563
+ },
564
+ {
565
+ "grad_norm": 0.7201863527297974,
566
+ "learning_rate": 9.119069424624165e-06,
567
+ "loss": 0.0487,
568
+ "step": 930
569
+ },
570
+ {
571
+ "grad_norm": 0.8865998387336731,
572
+ "learning_rate": 9.095496762130099e-06,
573
+ "loss": 0.0466,
574
+ "step": 940
575
+ },
576
+ {
577
+ "grad_norm": 1.0389326810836792,
578
+ "learning_rate": 9.071644177872594e-06,
579
+ "loss": 0.0576,
580
+ "step": 950
581
+ },
582
+ {
583
+ "grad_norm": 0.7415164709091187,
584
+ "learning_rate": 9.047513302144096e-06,
585
+ "loss": 0.0444,
586
+ "step": 960
587
+ },
588
+ {
589
+ "grad_norm": 0.7516604661941528,
590
+ "learning_rate": 9.023105784257906e-06,
591
+ "loss": 0.0467,
592
+ "step": 970
593
+ },
594
+ {
595
+ "grad_norm": 0.9304906725883484,
596
+ "learning_rate": 8.998423292435455e-06,
597
+ "loss": 0.0496,
598
+ "step": 980
599
+ },
600
+ {
601
+ "grad_norm": 0.8555199503898621,
602
+ "learning_rate": 8.973467513692265e-06,
603
+ "loss": 0.0527,
604
+ "step": 990
605
+ },
606
+ {
607
+ "grad_norm": 0.7662856578826904,
608
+ "learning_rate": 8.94824015372267e-06,
609
+ "loss": 0.0482,
610
+ "step": 1000
611
+ },
612
+ {
613
+ "grad_norm": 0.8362768292427063,
614
+ "learning_rate": 8.922742936783207e-06,
615
+ "loss": 0.0457,
616
+ "step": 1010
617
+ },
618
+ {
619
+ "grad_norm": 0.8270362615585327,
620
+ "learning_rate": 8.896977605574788e-06,
621
+ "loss": 0.0468,
622
+ "step": 1020
623
+ },
624
+ {
625
+ "grad_norm": 0.8668210506439209,
626
+ "learning_rate": 8.870945921123576e-06,
627
+ "loss": 0.0476,
628
+ "step": 1030
629
+ },
630
+ {
631
+ "grad_norm": 0.7505970597267151,
632
+ "learning_rate": 8.844649662660623e-06,
633
+ "loss": 0.0481,
634
+ "step": 1040
635
+ },
636
+ {
637
+ "grad_norm": 0.8698759078979492,
638
+ "learning_rate": 8.818090627500266e-06,
639
+ "loss": 0.0421,
640
+ "step": 1050
641
+ },
642
+ {
643
+ "grad_norm": 0.7640476226806641,
644
+ "learning_rate": 8.791270630917276e-06,
645
+ "loss": 0.0534,
646
+ "step": 1060
647
+ },
648
+ {
649
+ "grad_norm": 0.8142609000205994,
650
+ "learning_rate": 8.764191506022794e-06,
651
+ "loss": 0.0486,
652
+ "step": 1070
653
+ },
654
+ {
655
+ "grad_norm": 0.6571221947669983,
656
+ "learning_rate": 8.736855103639038e-06,
657
+ "loss": 0.0491,
658
+ "step": 1080
659
+ },
660
+ {
661
+ "grad_norm": 0.5830385684967041,
662
+ "learning_rate": 8.709263292172796e-06,
663
+ "loss": 0.0361,
664
+ "step": 1090
665
+ },
666
+ {
667
+ "grad_norm": 0.8238739371299744,
668
+ "learning_rate": 8.68141795748773e-06,
669
+ "loss": 0.043,
670
+ "step": 1100
671
+ },
672
+ {
673
+ "grad_norm": 0.8775807619094849,
674
+ "learning_rate": 8.65332100277548e-06,
675
+ "loss": 0.0481,
676
+ "step": 1110
677
+ },
678
+ {
679
+ "grad_norm": 0.7803472876548767,
680
+ "learning_rate": 8.624974348425575e-06,
681
+ "loss": 0.0465,
682
+ "step": 1120
683
+ },
684
+ {
685
+ "grad_norm": 0.8338977694511414,
686
+ "learning_rate": 8.596379931894187e-06,
687
+ "loss": 0.0501,
688
+ "step": 1130
689
+ },
690
+ {
691
+ "grad_norm": 0.6809723973274231,
692
+ "learning_rate": 8.567539707571704e-06,
693
+ "loss": 0.0493,
694
+ "step": 1140
695
+ },
696
+ {
697
+ "grad_norm": 0.779438316822052,
698
+ "learning_rate": 8.538455646649146e-06,
699
+ "loss": 0.0458,
700
+ "step": 1150
701
+ },
702
+ {
703
+ "grad_norm": 0.8775070309638977,
704
+ "learning_rate": 8.509129736983446e-06,
705
+ "loss": 0.044,
706
+ "step": 1160
707
+ },
708
+ {
709
+ "grad_norm": 0.738409698009491,
710
+ "learning_rate": 8.479563982961572e-06,
711
+ "loss": 0.042,
712
+ "step": 1170
713
+ },
714
+ {
715
+ "grad_norm": 0.7935025095939636,
716
+ "learning_rate": 8.449760405363538e-06,
717
+ "loss": 0.0433,
718
+ "step": 1180
719
+ },
720
+ {
721
+ "grad_norm": 0.9372588992118835,
722
+ "learning_rate": 8.419721041224286e-06,
723
+ "loss": 0.0446,
724
+ "step": 1190
725
+ },
726
+ {
727
+ "grad_norm": 0.7028743028640747,
728
+ "learning_rate": 8.389447943694451e-06,
729
+ "loss": 0.0456,
730
+ "step": 1200
731
+ },
732
+ {
733
+ "grad_norm": 0.7677087783813477,
734
+ "learning_rate": 8.358943181900032e-06,
735
+ "loss": 0.0439,
736
+ "step": 1210
737
+ },
738
+ {
739
+ "grad_norm": 0.6402320861816406,
740
+ "learning_rate": 8.328208840800981e-06,
741
+ "loss": 0.0374,
742
+ "step": 1220
743
+ },
744
+ {
745
+ "grad_norm": 0.9483605027198792,
746
+ "learning_rate": 8.297247021048685e-06,
747
+ "loss": 0.0438,
748
+ "step": 1230
749
+ },
750
+ {
751
+ "grad_norm": 0.710944652557373,
752
+ "learning_rate": 8.266059838842396e-06,
753
+ "loss": 0.0392,
754
+ "step": 1240
755
+ },
756
+ {
757
+ "grad_norm": 0.612249493598938,
758
+ "learning_rate": 8.23464942578459e-06,
759
+ "loss": 0.0454,
760
+ "step": 1250
761
+ },
762
+ {
763
+ "grad_norm": 0.6189194917678833,
764
+ "learning_rate": 8.203017928735278e-06,
765
+ "loss": 0.038,
766
+ "step": 1260
767
+ },
768
+ {
769
+ "grad_norm": 0.7650754451751709,
770
+ "learning_rate": 8.171167509665261e-06,
771
+ "loss": 0.0442,
772
+ "step": 1270
773
+ },
774
+ {
775
+ "grad_norm": 0.71224445104599,
776
+ "learning_rate": 8.139100345508378e-06,
777
+ "loss": 0.0419,
778
+ "step": 1280
779
+ },
780
+ {
781
+ "grad_norm": 0.9028244018554688,
782
+ "learning_rate": 8.106818628012698e-06,
783
+ "loss": 0.0446,
784
+ "step": 1290
785
+ },
786
+ {
787
+ "grad_norm": 0.746516227722168,
788
+ "learning_rate": 8.074324563590736e-06,
789
+ "loss": 0.042,
790
+ "step": 1300
791
+ },
792
+ {
793
+ "grad_norm": 0.6074661016464233,
794
+ "learning_rate": 8.041620373168628e-06,
795
+ "loss": 0.0386,
796
+ "step": 1310
797
+ },
798
+ {
799
+ "grad_norm": 0.7320516705513,
800
+ "learning_rate": 8.00870829203435e-06,
801
+ "loss": 0.0427,
802
+ "step": 1320
803
+ },
804
+ {
805
+ "grad_norm": 0.7389794588088989,
806
+ "learning_rate": 7.975590569684926e-06,
807
+ "loss": 0.0369,
808
+ "step": 1330
809
+ },
810
+ {
811
+ "grad_norm": 0.6718634366989136,
812
+ "learning_rate": 7.942269469672688e-06,
813
+ "loss": 0.0393,
814
+ "step": 1340
815
+ },
816
+ {
817
+ "grad_norm": 0.8955004811286926,
818
+ "learning_rate": 7.908747269450558e-06,
819
+ "loss": 0.0342,
820
+ "step": 1350
821
+ },
822
+ {
823
+ "grad_norm": 0.7186753153800964,
824
+ "learning_rate": 7.875026260216395e-06,
825
+ "loss": 0.0371,
826
+ "step": 1360
827
+ },
828
+ {
829
+ "grad_norm": 0.7176233530044556,
830
+ "learning_rate": 7.841108746756383e-06,
831
+ "loss": 0.03,
832
+ "step": 1370
833
+ },
834
+ {
835
+ "grad_norm": 0.689918577671051,
836
+ "learning_rate": 7.806997047287515e-06,
837
+ "loss": 0.0349,
838
+ "step": 1380
839
+ },
840
+ {
841
+ "grad_norm": 0.6836857199668884,
842
+ "learning_rate": 7.772693493299139e-06,
843
+ "loss": 0.0319,
844
+ "step": 1390
845
+ },
846
+ {
847
+ "grad_norm": 0.8493905067443848,
848
+ "learning_rate": 7.738200429393601e-06,
849
+ "loss": 0.0391,
850
+ "step": 1400
851
+ },
852
+ {
853
+ "grad_norm": 0.8006901741027832,
854
+ "learning_rate": 7.703520213126e-06,
855
+ "loss": 0.0317,
856
+ "step": 1410
857
+ },
858
+ {
859
+ "grad_norm": 0.7174612879753113,
860
+ "learning_rate": 7.66865521484305e-06,
861
+ "loss": 0.0389,
862
+ "step": 1420
863
+ },
864
+ {
865
+ "grad_norm": 0.6980575323104858,
866
+ "learning_rate": 7.633607817521074e-06,
867
+ "loss": 0.0357,
868
+ "step": 1430
869
+ },
870
+ {
871
+ "grad_norm": 0.7103762626647949,
872
+ "learning_rate": 7.59838041660312e-06,
873
+ "loss": 0.0362,
874
+ "step": 1440
875
+ },
876
+ {
877
+ "grad_norm": 0.7503710985183716,
878
+ "learning_rate": 7.5629754198352475e-06,
879
+ "loss": 0.0388,
880
+ "step": 1450
881
+ },
882
+ {
883
+ "grad_norm": 0.7302154898643494,
884
+ "learning_rate": 7.527395247101956e-06,
885
+ "loss": 0.0353,
886
+ "step": 1460
887
+ },
888
+ {
889
+ "grad_norm": 0.6966300010681152,
890
+ "learning_rate": 7.491642330260789e-06,
891
+ "loss": 0.0359,
892
+ "step": 1470
893
+ },
894
+ {
895
+ "grad_norm": 0.6958418488502502,
896
+ "learning_rate": 7.455719112976119e-06,
897
+ "loss": 0.0411,
898
+ "step": 1480
899
+ },
900
+ {
901
+ "grad_norm": 0.5858579874038696,
902
+ "learning_rate": 7.419628050552131e-06,
903
+ "loss": 0.0327,
904
+ "step": 1490
905
+ },
906
+ {
907
+ "grad_norm": 0.7254495620727539,
908
+ "learning_rate": 7.383371609764999e-06,
909
+ "loss": 0.0439,
910
+ "step": 1500
911
+ },
912
+ {
913
+ "grad_norm": 0.7627992630004883,
914
+ "learning_rate": 7.3469522686942885e-06,
915
+ "loss": 0.0343,
916
+ "step": 1510
917
+ },
918
+ {
919
+ "grad_norm": 0.7950798869132996,
920
+ "learning_rate": 7.310372516553585e-06,
921
+ "loss": 0.035,
922
+ "step": 1520
923
+ },
924
+ {
925
+ "grad_norm": 0.683579683303833,
926
+ "learning_rate": 7.273634853520357e-06,
927
+ "loss": 0.0362,
928
+ "step": 1530
929
+ },
930
+ {
931
+ "grad_norm": 0.6747159361839294,
932
+ "learning_rate": 7.2367417905650725e-06,
933
+ "loss": 0.0413,
934
+ "step": 1540
935
+ },
936
+ {
937
+ "grad_norm": 0.7270603179931641,
938
+ "learning_rate": 7.199695849279576e-06,
939
+ "loss": 0.037,
940
+ "step": 1550
941
+ },
942
+ {
943
+ "grad_norm": 0.6702049970626831,
944
+ "learning_rate": 7.1624995617047475e-06,
945
+ "loss": 0.0334,
946
+ "step": 1560
947
+ },
948
+ {
949
+ "grad_norm": 0.8103553056716919,
950
+ "learning_rate": 7.12515547015743e-06,
951
+ "loss": 0.0321,
952
+ "step": 1570
953
+ },
954
+ {
955
+ "grad_norm": 0.7140473127365112,
956
+ "learning_rate": 7.0876661270566745e-06,
957
+ "loss": 0.032,
958
+ "step": 1580
959
+ },
960
+ {
961
+ "grad_norm": 0.6398239135742188,
962
+ "learning_rate": 7.050034094749286e-06,
963
+ "loss": 0.0345,
964
+ "step": 1590
965
+ },
966
+ {
967
+ "grad_norm": 0.7120059132575989,
968
+ "learning_rate": 7.012261945334683e-06,
969
+ "loss": 0.0325,
970
+ "step": 1600
971
+ },
972
+ {
973
+ "grad_norm": 0.6111170053482056,
974
+ "learning_rate": 6.974352260489104e-06,
975
+ "loss": 0.0291,
976
+ "step": 1610
977
+ },
978
+ {
979
+ "grad_norm": 0.6516941785812378,
980
+ "learning_rate": 6.936307631289148e-06,
981
+ "loss": 0.0278,
982
+ "step": 1620
983
+ },
984
+ {
985
+ "grad_norm": 0.686859130859375,
986
+ "learning_rate": 6.898130658034685e-06,
987
+ "loss": 0.0291,
988
+ "step": 1630
989
+ },
990
+ {
991
+ "grad_norm": 0.8052101135253906,
992
+ "learning_rate": 6.859823950071127e-06,
993
+ "loss": 0.0324,
994
+ "step": 1640
995
+ },
996
+ {
997
+ "grad_norm": 0.7104664444923401,
998
+ "learning_rate": 6.8213901256110785e-06,
999
+ "loss": 0.036,
1000
+ "step": 1650
1001
+ },
1002
+ {
1003
+ "grad_norm": 0.5812593698501587,
1004
+ "learning_rate": 6.782831811555386e-06,
1005
+ "loss": 0.0252,
1006
+ "step": 1660
1007
+ },
1008
+ {
1009
+ "grad_norm": 0.7957876920700073,
1010
+ "learning_rate": 6.744151643313597e-06,
1011
+ "loss": 0.0318,
1012
+ "step": 1670
1013
+ },
1014
+ {
1015
+ "grad_norm": 0.6455429196357727,
1016
+ "learning_rate": 6.705352264623829e-06,
1017
+ "loss": 0.0337,
1018
+ "step": 1680
1019
+ },
1020
+ {
1021
+ "grad_norm": 0.6577746272087097,
1022
+ "learning_rate": 6.666436327372079e-06,
1023
+ "loss": 0.0352,
1024
+ "step": 1690
1025
+ },
1026
+ {
1027
+ "grad_norm": 0.7138893008232117,
1028
+ "learning_rate": 6.627406491410961e-06,
1029
+ "loss": 0.0312,
1030
+ "step": 1700
1031
+ },
1032
+ {
1033
+ "grad_norm": 0.6545776724815369,
1034
+ "learning_rate": 6.58826542437792e-06,
1035
+ "loss": 0.0318,
1036
+ "step": 1710
1037
+ },
1038
+ {
1039
+ "grad_norm": 0.5184334516525269,
1040
+ "learning_rate": 6.549015801512895e-06,
1041
+ "loss": 0.0349,
1042
+ "step": 1720
1043
+ },
1044
+ {
1045
+ "grad_norm": 0.6423726081848145,
1046
+ "learning_rate": 6.509660305475468e-06,
1047
+ "loss": 0.0316,
1048
+ "step": 1730
1049
+ },
1050
+ {
1051
+ "grad_norm": 0.5985066890716553,
1052
+ "learning_rate": 6.47020162616152e-06,
1053
+ "loss": 0.0279,
1054
+ "step": 1740
1055
+ },
1056
+ {
1057
+ "grad_norm": 0.7095875144004822,
1058
+ "learning_rate": 6.430642460519365e-06,
1059
+ "loss": 0.0314,
1060
+ "step": 1750
1061
+ },
1062
+ {
1063
+ "grad_norm": 0.7356165051460266,
1064
+ "learning_rate": 6.390985512365426e-06,
1065
+ "loss": 0.0351,
1066
+ "step": 1760
1067
+ },
1068
+ {
1069
+ "grad_norm": 0.6364956498146057,
1070
+ "learning_rate": 6.351233492199432e-06,
1071
+ "loss": 0.0327,
1072
+ "step": 1770
1073
+ },
1074
+ {
1075
+ "grad_norm": 0.6260037422180176,
1076
+ "learning_rate": 6.311389117019155e-06,
1077
+ "loss": 0.0295,
1078
+ "step": 1780
1079
+ },
1080
+ {
1081
+ "grad_norm": 0.770182192325592,
1082
+ "learning_rate": 6.271455110134713e-06,
1083
+ "loss": 0.0368,
1084
+ "step": 1790
1085
+ },
1086
+ {
1087
+ "grad_norm": 0.6314713954925537,
1088
+ "learning_rate": 6.231434200982429e-06,
1089
+ "loss": 0.0309,
1090
+ "step": 1800
1091
+ },
1092
+ {
1093
+ "grad_norm": 0.5799538493156433,
1094
+ "learning_rate": 6.191329124938284e-06,
1095
+ "loss": 0.0303,
1096
+ "step": 1810
1097
+ },
1098
+ {
1099
+ "grad_norm": 0.6934332251548767,
1100
+ "learning_rate": 6.15114262313095e-06,
1101
+ "loss": 0.0274,
1102
+ "step": 1820
1103
+ },
1104
+ {
1105
+ "grad_norm": 0.5117380023002625,
1106
+ "learning_rate": 6.110877442254445e-06,
1107
+ "loss": 0.0276,
1108
+ "step": 1830
1109
+ },
1110
+ {
1111
+ "grad_norm": 0.5944725275039673,
1112
+ "learning_rate": 6.0705363343803945e-06,
1113
+ "loss": 0.0272,
1114
+ "step": 1840
1115
+ },
1116
+ {
1117
+ "grad_norm": 0.7068486213684082,
1118
+ "learning_rate": 6.030122056769935e-06,
1119
+ "loss": 0.0271,
1120
+ "step": 1850
1121
+ },
1122
+ {
1123
+ "grad_norm": 0.6229940056800842,
1124
+ "learning_rate": 5.989637371685258e-06,
1125
+ "loss": 0.032,
1126
+ "step": 1860
1127
+ },
1128
+ {
1129
+ "grad_norm": 0.5604724884033203,
1130
+ "learning_rate": 5.949085046200808e-06,
1131
+ "loss": 0.0319,
1132
+ "step": 1870
1133
+ },
1134
+ {
1135
+ "grad_norm": 0.6178894639015198,
1136
+ "learning_rate": 5.908467852014169e-06,
1137
+ "loss": 0.0258,
1138
+ "step": 1880
1139
+ },
1140
+ {
1141
+ "grad_norm": 0.5318469405174255,
1142
+ "learning_rate": 5.867788565256607e-06,
1143
+ "loss": 0.0264,
1144
+ "step": 1890
1145
+ },
1146
+ {
1147
+ "grad_norm": 0.5695937871932983,
1148
+ "learning_rate": 5.827049966303336e-06,
1149
+ "loss": 0.0361,
1150
+ "step": 1900
1151
+ },
1152
+ {
1153
+ "grad_norm": 0.6509976983070374,
1154
+ "learning_rate": 5.786254839583478e-06,
1155
+ "loss": 0.0313,
1156
+ "step": 1910
1157
+ },
1158
+ {
1159
+ "grad_norm": 0.8354060053825378,
1160
+ "learning_rate": 5.745405973389757e-06,
1161
+ "loss": 0.0287,
1162
+ "step": 1920
1163
+ },
1164
+ {
1165
+ "grad_norm": 0.6091945171356201,
1166
+ "learning_rate": 5.704506159687914e-06,
1167
+ "loss": 0.0297,
1168
+ "step": 1930
1169
+ },
1170
+ {
1171
+ "grad_norm": 0.6996727585792542,
1172
+ "learning_rate": 5.6635581939258855e-06,
1173
+ "loss": 0.0282,
1174
+ "step": 1940
1175
+ },
1176
+ {
1177
+ "grad_norm": 0.6515761613845825,
1178
+ "learning_rate": 5.622564874842742e-06,
1179
+ "loss": 0.0288,
1180
+ "step": 1950
1181
+ },
1182
+ {
1183
+ "grad_norm": 0.6472054719924927,
1184
+ "learning_rate": 5.581529004277384e-06,
1185
+ "loss": 0.0232,
1186
+ "step": 1960
1187
+ },
1188
+ {
1189
+ "grad_norm": 0.6369526386260986,
1190
+ "learning_rate": 5.540453386977058e-06,
1191
+ "loss": 0.0294,
1192
+ "step": 1970
1193
+ },
1194
+ {
1195
+ "grad_norm": 0.6562472581863403,
1196
+ "learning_rate": 5.4993408304056425e-06,
1197
+ "loss": 0.0357,
1198
+ "step": 1980
1199
+ },
1200
+ {
1201
+ "grad_norm": 0.7148199677467346,
1202
+ "learning_rate": 5.458194144551769e-06,
1203
+ "loss": 0.0267,
1204
+ "step": 1990
1205
+ },
1206
+ {
1207
+ "grad_norm": 0.5509341955184937,
1208
+ "learning_rate": 5.417016141736756e-06,
1209
+ "loss": 0.0241,
1210
+ "step": 2000
1211
+ }
1212
+ ],
1213
+ "logging_steps": 10,
1214
+ "max_steps": 4000,
1215
+ "num_input_tokens_seen": 0,
1216
+ "num_train_epochs": 9223372036854775807,
1217
+ "save_steps": 500,
1218
+ "stateful_callbacks": {
1219
+ "TrainerControl": {
1220
+ "args": {
1221
+ "should_epoch_stop": false,
1222
+ "should_evaluate": false,
1223
+ "should_log": false,
1224
+ "should_save": true,
1225
+ "should_training_stop": false
1226
+ },
1227
+ "attributes": {}
1228
+ }
1229
+ },
1230
+ "total_flos": 0.0,
1231
+ "train_batch_size": 32,
1232
+ "trial_name": null,
1233
+ "trial_params": null
1234
+ }
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:78d72d3f0ccfbe2c2d1ec0dbeec3ec4e9e481448654f386ff97b7edee0243b49
3
+ size 8017
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/training_plan.json ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "a18cr2",
3
+ "frames": 66850,
4
+ "source": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/groot-n17-base",
5
+ "recipe": "full fine-tune (tune_llm, tune_visual, projector, diffusion, vlln), fixed steps, 2 nodes x 8 x 32, no accumulation; variant 2026-11-15: all dropouts 0, action horizon 32, 16 flow sampling steps, layer-wise lr",
6
+ "model_overrides": {
7
+ "attn_dropout": 0.0,
8
+ "diffusion_model_cfg.dropout": 0.0,
9
+ "state_dropout_prob": 0.0,
10
+ "num_inference_timesteps": 16
11
+ },
12
+ "lr_groups": {
13
+ "language": 1e-06,
14
+ "vision": 1e-05,
15
+ "projector": 0.0001,
16
+ "rest": 0.0001
17
+ },
18
+ "per_gpu_batch": 32,
19
+ "gradient_accumulation_steps": 1,
20
+ "num_nodes": 2,
21
+ "world_size": 16,
22
+ "global_batch": 512,
23
+ "max_steps": 4000,
24
+ "learning_rate": 0.0001,
25
+ "warmup_ratio": 0.05,
26
+ "lr_schedule": "linear warmup 5 %, cosine to 0",
27
+ "save_steps": 500,
28
+ "save_list": [
29
+ 2000,
30
+ 2500,
31
+ 3000,
32
+ 3500,
33
+ 4000
34
+ ],
35
+ "epochs_equivalent": 30.64
36
+ }
a18-pour-water/b512-2cam-h32-cropped-v2/step-002000/wandb_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"project": "finetune-gr00t-n1d7", "run_id": "production"}
a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/cache_cosmos_reason2.py ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Put the full gated nvidia/Cosmos-Reason2-2B (pinned 9ce19a1) into the shared HF hub cache $G/hf-home/hub.
2
+
3
+ GR00T N1.7 builds its backbone with Qwen3VLForConditionalGeneration.from_pretrained("nvidia/Cosmos-Reason2-2B") and
4
+ its processor with Qwen3VLProcessor.from_pretrained(same name) before loading the GR00T checkpoint, so every node needs
5
+ the whole repo resolvable by name. groot_node.sh runs with HF_HUB_OFFLINE=1, so nodes read this cache and never need
6
+ the token. refs/main is pinned to the verified revision.
7
+
8
+ Usage (token from stdin, kept in-process only):
9
+ $G/envs/groot-py312/bin/python cache_cosmos_reason2.py $G/hf-home/hub < /root/.hb-secrets/hf_token
10
+ """
11
+ import hashlib
12
+ import json
13
+ import os
14
+ import sys
15
+ import time
16
+
17
+ from huggingface_hub import HfApi, snapshot_download
18
+
19
+ REPO = "nvidia/Cosmos-Reason2-2B"
20
+ REV = "9ce19a195e423419c349abfc86fd07178b230561"
21
+ cache = os.path.abspath(sys.argv[1])
22
+ token = sys.stdin.readline().strip()
23
+ if not token.startswith("hf_"):
24
+ raise SystemExit("token format rejected")
25
+ api = HfApi(endpoint="https://huggingface.co", token=token)
26
+ info = api.model_info(REPO, revision=REV, files_metadata=True)
27
+ assert info.sha == REV, info.sha
28
+ t0 = time.monotonic()
29
+ for attempt in range(1, 21):
30
+ try:
31
+ path = snapshot_download(REPO, revision=REV, cache_dir=cache, token=token, max_workers=8)
32
+ break
33
+ except Exception as error: # transient stalls
34
+ if any(k in str(error).lower() for k in ("gated", "401", "403")):
35
+ raise
36
+ print(f"retry {attempt}: {type(error).__name__}: {str(error)[:200]}", flush=True)
37
+ time.sleep(10)
38
+ else:
39
+ raise SystemExit("download did not finish")
40
+ token = None
41
+ count, size = 0, 0
42
+ for s in info.siblings:
43
+ f = os.path.join(path, s.rfilename)
44
+ assert os.path.isfile(f), s.rfilename
45
+ if s.size is not None:
46
+ assert os.path.getsize(f) == s.size, s.rfilename
47
+ if s.lfs is not None:
48
+ h = hashlib.sha256()
49
+ with open(f, "rb") as stream:
50
+ for chunk in iter(lambda: stream.read(1 << 22), b""):
51
+ h.update(chunk)
52
+ assert h.hexdigest() == s.lfs.sha256, s.rfilename
53
+ count += 1
54
+ size += s.size or 0
55
+ refs = os.path.join(cache, "models--" + REPO.replace("/", "--"), "refs")
56
+ os.makedirs(refs, exist_ok=True)
57
+ with open(os.path.join(refs, "main"), "w") as f:
58
+ f.write(REV)
59
+ print("CACHE_OK", json.dumps({"repo": REPO, "revision": REV, "snapshot": path, "files": count, "bytes": size,
60
+ "seconds": round(time.monotonic() - t0, 1)}), flush=True)
a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/check_reason2_offline.py ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """CPU check that GR00T's Cosmos-Reason2-2B pieces load offline from $G/hf-home exactly as groot_node.sh will.
2
+
3
+ Usage: HF_HOME=$G/hf-home HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 $G/envs/groot-py312/bin/python check_reason2_offline.py
4
+ """
5
+ import json
6
+ import os
7
+ import sys
8
+ from pathlib import Path
9
+
10
+ import torch
11
+
12
+ sys.path.insert(0, str(Path(__file__).resolve().parent))
13
+ import hb_offline_fix # noqa: E402,F401
14
+ from transformers import AutoConfig, Qwen3VLForConditionalGeneration, Qwen3VLProcessor
15
+
16
+ assert os.environ.get("HF_HUB_OFFLINE") == "1", "run with HF_HUB_OFFLINE=1"
17
+ name = "nvidia/Cosmos-Reason2-2B"
18
+ torch.set_num_threads(8)
19
+ proc = Qwen3VLProcessor.from_pretrained(name, trust_remote_code=True)
20
+ cfg = AutoConfig.from_pretrained(name, trust_remote_code=True)
21
+ ids = proc.tokenizer("pick up the trash and drop it in the bin")["input_ids"]
22
+ model = Qwen3VLForConditionalGeneration.from_pretrained(name, torch_dtype=torch.bfloat16)
23
+ n = sum(p.numel() for p in model.parameters())
24
+ print("REASON2_OFFLINE_OK", json.dumps({"processor": type(proc).__name__, "tokens": len(ids),
25
+ "model_type": cfg.model_type, "params_M": round(n / 1e6, 1),
26
+ "llm_layers": len(model.model.language_model.layers)}), flush=True)
a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/g1_feedback76_h16.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """G1 feedback76 / H16 GR00T modality config (copy of HumanWAM configs/humanwam/groot/feedback76_h16.py @ c76527b).
2
+
3
+ The matching modality.json is generated per dataset from the recorded joint names (prepare_groot_data.py);
4
+ HumanWAM's feedback76_modality.json has right_arm at observation.state[29:36] (left-hand fingers), which is wrong
5
+ for our G1 exports (right arm is [22:29]).
6
+ """
7
+ from gr00t.configs.data.embodiment_configs import register_modality_config
8
+ from gr00t.data.embodiment_tags import EmbodimentTag
9
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
10
+
11
+ feedback76_h16 = {
12
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["ego_view", "chest"]),
13
+ "state": ModalityConfig(
14
+ delta_indices=[0],
15
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "left_hand", "right_hand",
16
+ "projected_gravity"],
17
+ ),
18
+ "action": ModalityConfig(
19
+ delta_indices=list(range(16)),
20
+ modality_keys=["motion_token", "left_hand_joints", "right_hand_joints"],
21
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
22
+ format=ActionFormat.DEFAULT) for _ in range(3)],
23
+ ),
24
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
25
+ }
26
+
27
+ register_modality_config(feedback76_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/groot_chain.py ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Per-node GR00T relay: per-layout batch-size probe (largest that fits), then 20-epoch production per task.
2
+
3
+ Run on every node (nohup), node rank from the DLC env RANK: python groot_chain.py [task ...]
4
+ Default tasks: b6 a3 a7 a18 (basket postponed). A phase succeeds only if all 4 node launchers exit 0.
5
+
6
+ Batch (user rule: as large as memory allows, 2026-11-10): descending CANDIDATES, 4 steps each on all 32 GPUs; a
7
+ candidate fits when all launchers exit 0 AND the peak reserved memory of every rank (written by hb_launch.py to
8
+ logs/peak-<phase>-rank<r>.json) is <= MAX_FRAC of the GPU. Node 0 alone judges each probe and writes a verdict file
9
+ the other nodes wait for, so all nodes take the same decision. One decision per robot layout (g1 / h2).
10
+ Outputs: $G/groot-runs-bmax/ (the earlier per-GPU-64 runs stay in $G/groot-runs/).
11
+ """
12
+ import fcntl
13
+ import json
14
+ import math
15
+ import os
16
+ import shutil
17
+ import subprocess
18
+ import sys
19
+ import time
20
+ from pathlib import Path
21
+
22
+ H = Path(__file__).resolve().parent
23
+ G = H.parent
24
+ RUNS = G / os.environ.get("HB_GROOT_RUNS_NAME", "groot-runs-bmax")
25
+ NODE = int(os.environ["RANK"])
26
+ TASKS = sys.argv[1:] or ["b6", "a3", "a7", "a18"] # basket postponed (user plan 2026-11-10)
27
+ CANDIDATES = [320, 288, 256, 224, 192, 160, 128, 96, 64, 48, 32]
28
+ MAX_FRAC = 0.92
29
+ EPOCHS = 20
30
+ PORT_BASE = int(os.environ.get("HB_GROOT_PORT_BASE", "32500")) # probe PORT_BASE+i, production PORT_BASE+50+k
31
+ RUNS.mkdir(parents=True, exist_ok=True)
32
+ os.environ["HB_GROOT_RUNS"] = str(RUNS) # read by groot_node.sh
33
+ lock = (RUNS / f"chain-node-{NODE}.lock").open("a")
34
+ fcntl.flock(lock, fcntl.LOCK_EX | fcntl.LOCK_NB)
35
+
36
+
37
+ def log(msg):
38
+ line = f"{time.strftime('%Y-%m-%d %H:%M:%S')} node{NODE} {msg}"
39
+ print(line, flush=True)
40
+ with (RUNS / f"chain-node-{NODE}.log").open("a") as f:
41
+ f.write(line + "\n")
42
+
43
+
44
+ def run_phase(task, phase, bs, steps, save, port):
45
+ logs = RUNS / task / "logs"
46
+ exits = [logs / f"{phase}-node-{r}.exit" for r in range(4)]
47
+ if all(p.is_file() for p in exits):
48
+ return all(p.read_text().strip() == "0" for p in exits)
49
+ subprocess.run(["bash", str(H / "groot_node.sh"), task, phase, str(bs), str(steps), str(save), str(port)])
50
+ deadline = time.monotonic() + 600
51
+ while not all(p.is_file() for p in exits):
52
+ if time.monotonic() > deadline:
53
+ log(f"{task}/{phase}: timed out waiting for peer exit files")
54
+ return False
55
+ time.sleep(5)
56
+ ok = all(p.read_text().strip() == "0" for p in exits)
57
+ log(f"{task}/{phase}: bs={bs} steps={steps} -> {'OK' if ok else 'FAIL'}")
58
+ return ok
59
+
60
+
61
+ def peak_frac(task, phase):
62
+ """Max over ranks of peak reserved / total; None when no rank reported."""
63
+ fracs = []
64
+ for f in (RUNS / task / "logs").glob(f"peak-{phase}-rank*.json"):
65
+ try:
66
+ d = json.loads(f.read_text())
67
+ fracs.append(d["peak_reserved_gb"] / d["total_gb"])
68
+ except (ValueError, KeyError, OSError):
69
+ continue
70
+ return (max(fracs), len(fracs)) if fracs else (None, 0)
71
+
72
+
73
+ def probe_verdict(task, cand, port):
74
+ phase = f"probe-bs{cand}"
75
+ verdict = RUNS / task / "logs" / f"{phase}.verdict.json"
76
+ ok = run_phase(task, phase, cand, 4, 1000000, port)
77
+ if NODE == 0 and not verdict.is_file():
78
+ time.sleep(10) # let peak files from the other nodes become visible on NAS
79
+ frac, n = peak_frac(task, phase) if ok else (None, 0)
80
+ fits = bool(ok and frac is not None and frac <= MAX_FRAC)
81
+ verdict.write_text(json.dumps({"per_gpu_batch": cand, "ran_ok": ok, "peak_frac": frac, "ranks_reported": n,
82
+ "max_frac": MAX_FRAC, "fits": fits}) + "\n")
83
+ deadline = time.monotonic() + 600
84
+ while not verdict.is_file():
85
+ if time.monotonic() > deadline:
86
+ log(f"{task}/{phase}: no verdict from node 0")
87
+ return False
88
+ time.sleep(5)
89
+ v = json.loads(verdict.read_text())
90
+ log(f"{task}/{phase}: verdict {v}")
91
+ return v["fits"]
92
+
93
+
94
+ def prune_resume_state(out):
95
+ """Keep weights of every checkpoint but resume state (DeepSpeed optimizer shards, scheduler, RNG) only in the latest."""
96
+ ckpts = sorted((p for p in out.glob("checkpoint-*") if p.name.split("-")[-1].isdigit()),
97
+ key=lambda p: int(p.name.split("-")[-1]))
98
+ for c in ckpts[:-1]:
99
+ for p in list(c.glob("global_step*")) + list(c.glob("optimizer*.pt")) + list(c.glob("scheduler.pt")) \
100
+ + list(c.glob("rng_state*.pth")) + list(c.glob("latest")) + list(c.glob("zero_to_fp32.py")):
101
+ if p.is_dir():
102
+ shutil.rmtree(p, ignore_errors=True)
103
+ else:
104
+ p.unlink(missing_ok=True)
105
+ log(f"{c.name}: resume state dropped (weights kept)")
106
+
107
+
108
+ def batch_for(task):
109
+ layout = "h2" if task in ("a17", "a18") else "g1"
110
+ decision = RUNS / f"batch-decision-{layout}.json"
111
+ if decision.is_file():
112
+ return json.loads(decision.read_text())["per_gpu_batch"]
113
+ for i, cand in enumerate(CANDIDATES):
114
+ if probe_verdict(task, cand, PORT_BASE + i + (20 if layout == "h2" else 0)):
115
+ if NODE == 0:
116
+ decision.write_text(json.dumps({"per_gpu_batch": cand, "global_batch": cand * 32, "probe_task": task,
117
+ "layout": layout, "candidates": CANDIDATES,
118
+ "max_frac": MAX_FRAC}) + "\n")
119
+ return cand
120
+ time.sleep(20)
121
+ log(f"{layout}: no batch size fit; stopping")
122
+ sys.exit(1)
123
+
124
+
125
+ for k, task in enumerate(TASKS):
126
+ bs = batch_for(task)
127
+ log(f"{task}: per-GPU batch {bs} (global {bs * 32})")
128
+ receipt = json.loads((G / "groot-data" / task / "hb-dataset-receipt.json").read_text())
129
+ steps = math.ceil(EPOCHS * receipt["frames"] / (bs * 32))
130
+ save = math.ceil(steps / 2)
131
+ if not run_phase(task, "production", bs, steps, save, PORT_BASE + 50 + k):
132
+ log(f"{task}: production failed; stopping chain")
133
+ sys.exit(1)
134
+ if NODE == 0:
135
+ prune_resume_state(RUNS / task / "production")
136
+ (RUNS / task / "production-complete.json").write_text(json.dumps(
137
+ {"task": task, "frames": receipt["frames"], "epochs": EPOCHS, "per_gpu_batch": bs,
138
+ "global_batch": bs * 32, "max_steps": steps, "save_steps": save}) + "\n")
139
+ log("all tasks done")
a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/h2_body31_hand44_h16.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A18 H2 (Sharpa hands) GR00T modality config, same protocol as the A18 Delta-flow run
2
+ (schema a18-h2-sharpa-body31-hand44-state78-chunk16-v1): cameras head+chest, state = 75 joints + projected
3
+ gravity, action = action.body_target (31) + action.hands (44), horizon 16, absolute targets."""
4
+ from gr00t.configs.data.embodiment_configs import register_modality_config
5
+ from gr00t.data.embodiment_tags import EmbodimentTag
6
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
7
+
8
+ h2_body31_hand44_h16 = {
9
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["head", "chest"]),
10
+ "state": ModalityConfig(
11
+ delta_indices=[0],
12
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "head", "left_hand", "right_hand",
13
+ "projected_gravity"],
14
+ ),
15
+ "action": ModalityConfig(
16
+ delta_indices=list(range(16)),
17
+ modality_keys=["body_target", "left_hand_target", "right_hand_target"],
18
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
19
+ format=ActionFormat.DEFAULT) for _ in range(3)],
20
+ ),
21
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
22
+ }
23
+
24
+ register_modality_config(h2_body31_hand44_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/h2_body31_hand44_h32.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """2026-11-15 user: action horizon 32 copy of h2_body31_hand44_h16.py. A18 H2 (Sharpa hands) GR00T modality config, same protocol as the A18 Delta-flow run
2
+ (schema a18-h2-sharpa-body31-hand44-state78-chunk16-v1): cameras head+chest, state = 75 joints + projected
3
+ gravity, action = action.body_target (31) + action.hands (44), horizon 16, absolute targets."""
4
+ from gr00t.configs.data.embodiment_configs import register_modality_config
5
+ from gr00t.data.embodiment_tags import EmbodimentTag
6
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
7
+
8
+ h2_body31_hand44_h32 = {
9
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["head", "chest"]),
10
+ "state": ModalityConfig(
11
+ delta_indices=[0],
12
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "head", "left_hand", "right_hand",
13
+ "projected_gravity"],
14
+ ),
15
+ "action": ModalityConfig(
16
+ delta_indices=list(range(32)),
17
+ modality_keys=["body_target", "left_hand_target", "right_hand_target"],
18
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
19
+ format=ActionFormat.DEFAULT) for _ in range(3)],
20
+ ),
21
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
22
+ }
23
+
24
+ register_modality_config(h2_body31_hand44_h32, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/h2_body31_hand44_planner7_h16.py ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A14 v2 H2 (Sharpa hands) GR00T modality config (H2 training guide hb_training/docs/h2/a14_training_guide.md 4.1):
2
+ cameras head + chest, state78 = 75 joints + projected gravity, action82 = action.body_target (31) + action.hands (44)
3
+ + action.planner (7, the SONIC planner command), horizon 16, absolute targets. Identical to h2_body31_hand44_h16.py
4
+ except the added `planner` action modality."""
5
+ from gr00t.configs.data.embodiment_configs import register_modality_config
6
+ from gr00t.data.embodiment_tags import EmbodimentTag
7
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
8
+
9
+ h2_body31_hand44_planner7_h16 = {
10
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["head", "chest"]),
11
+ "state": ModalityConfig(
12
+ delta_indices=[0],
13
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "head", "left_hand", "right_hand",
14
+ "projected_gravity"],
15
+ ),
16
+ "action": ModalityConfig(
17
+ delta_indices=list(range(16)),
18
+ modality_keys=["body_target", "left_hand_target", "right_hand_target", "planner"],
19
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
20
+ format=ActionFormat.DEFAULT) for _ in range(4)],
21
+ ),
22
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
23
+ }
24
+
25
+ register_modality_config(h2_body31_hand44_planner7_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/h2w3_body31_hand44_h16.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A14 H2 (Sharpa hands) GR00T modality config with THREE cameras head + chest + wrist (user 2026-11-12; the third-person camera is never a model input). Otherwise identical to h2_body31_hand44_h16.py:, same protocol as the A18 Delta-flow run
2
+ (schema a18-h2-sharpa-body31-hand44-state78-chunk16-v1): cameras head+chest, state = 75 joints + projected
3
+ gravity, action = action.body_target (31) + action.hands (44), horizon 16, absolute targets."""
4
+ from gr00t.configs.data.embodiment_configs import register_modality_config
5
+ from gr00t.data.embodiment_tags import EmbodimentTag
6
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
7
+
8
+ h2w3_body31_hand44_h16 = {
9
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["head", "chest", "wrist"]),
10
+ "state": ModalityConfig(
11
+ delta_indices=[0],
12
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "head", "left_hand", "right_hand",
13
+ "projected_gravity"],
14
+ ),
15
+ "action": ModalityConfig(
16
+ delta_indices=list(range(16)),
17
+ modality_keys=["body_target", "left_hand_target", "right_hand_target"],
18
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
19
+ format=ActionFormat.DEFAULT) for _ in range(3)],
20
+ ),
21
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
22
+ }
23
+
24
+ register_modality_config(h2w3_body31_hand44_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/hb_aug_v2.py ADDED
@@ -0,0 +1,163 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Humanoid Benchmark data augmentation v2 (project CLAUDE.md), model-agnostic torch implementation.
2
+
3
+ augment(x, rng, obs_mask=None) takes x: float [T, V, 3, H, W] in [0, 1] (T frames, V cameras) and returns the
4
+ augmented tensor. Each call is one training sample: one of the 15 effects is chosen uniformly and its magnitude
5
+ is scaled by 1.5 around identity; the rest stay at x1. Colour is shared across cameras and time; geometry, light
6
+ patches and white balance are drawn per camera and fixed over time; sensor/optics effects (noise, gaussian blur,
7
+ JPEG, motion blur, sharpen, defocus) apply only to observation frames (obs_mask True).
8
+ Pixel-size parameters are defined at 224 px and scaled by H / 224.
9
+ perturb_joints(q, rng) applies the proprio spec: per-joint bias +-2 deg plus N(0, 0.25 deg) clipped at 3 sigma.
10
+ """
11
+ import io
12
+ import math
13
+
14
+ import numpy as np
15
+ import torch
16
+ import torch.nn.functional as F
17
+ from PIL import Image
18
+ from torchvision.transforms import InterpolationMode
19
+ from torchvision.transforms import functional as TF
20
+
21
+ EFFECTS = ("exposure", "contrast", "saturation", "hue", "gamma", "temperature", "rotation_crop", "noise",
22
+ "gaussian_blur", "jpeg", "motion_blur", "sharpen", "defocus", "light_shadow", "wb_channel")
23
+ CENTER = {"exposure": 0.0, "contrast": 1.0, "saturation": 1.0, "hue": 0.0, "gamma": 1.0, "temperature": 0.0}
24
+ RANGE = {"exposure": (-0.22, 0.22), "contrast": (0.88, 1.12), "saturation": (0.76, 1.24), "hue": (-0.015, 0.015),
25
+ "gamma": (0.89, 1.11), "temperature": (-0.066, 0.066)}
26
+ VERSION = "hb-aug-v2-20261109"
27
+
28
+
29
+ def sample(rng, views, height):
30
+ boost = EFFECTS[int(rng.integers(len(EFFECTS)))]
31
+
32
+ def k(name):
33
+ return 1.5 if boost == name else 1.0
34
+
35
+ px = height / 224.0
36
+ p = {"boost": boost}
37
+ for name, (a, b) in RANGE.items():
38
+ c = CENTER[name]
39
+ p[name] = float(rng.uniform(c + k(name) * (a - c), c + k(name) * (b - c)))
40
+ p["geometry"] = []
41
+ for _ in range(views):
42
+ angle = float(rng.uniform(-1.5 * k("rotation_crop"), 1.5 * k("rotation_crop")))
43
+ p["geometry"].append((angle, 1.0 + abs(angle) * 0.015 + float(rng.uniform(0, 0.02 * k("rotation_crop")))))
44
+ p["noise"] = 4.0 / 255.0 * k("noise") if rng.random() < 0.5 else 0.0
45
+ p["gaussian_blur"] = float(rng.uniform(0.25, 0.65)) * k("gaussian_blur") if rng.random() < 0.15 else 0.0
46
+ if rng.random() < 0.35:
47
+ lo = int(round(100 - k("jpeg") * 30))
48
+ hi = int(round(100 - k("jpeg") * 5))
49
+ p["jpeg"] = int(rng.integers(lo, hi + 1))
50
+ else:
51
+ p["jpeg"] = None
52
+ p["motion_blur"] = ((float(rng.uniform(3, 7)) * k("motion_blur") * px, float(rng.uniform(0, 180)))
53
+ if rng.random() < 0.15 else None)
54
+ p["sharpen"] = float(rng.uniform(0.2, 0.6)) * k("sharpen") if rng.random() < 0.15 else 0.0
55
+ p["defocus"] = float(rng.uniform(1, 2)) * k("defocus") * px if rng.random() < 0.10 else 0.0
56
+ if rng.random() < 0.30:
57
+ g = 0.30 * k("light_shadow")
58
+ p["light_shadow"] = [[(float(rng.uniform(0.1, 0.9)), float(rng.uniform(0.1, 0.9)),
59
+ float(rng.uniform(0.15, 0.35)), float(rng.uniform(1 - g, 1 + g)))
60
+ for _ in range(int(rng.integers(1, 3)))] for _ in range(views)]
61
+ else:
62
+ p["light_shadow"] = None
63
+ if rng.random() < 0.30:
64
+ g = 0.10 * k("wb_channel")
65
+ gains = []
66
+ for _ in range(views):
67
+ row = [1.0, 1.0, 1.0]
68
+ row[int(rng.integers(0, 3))] = float(rng.uniform(1 - g, 1 + g))
69
+ gains.append(row)
70
+ p["wb_channel"] = gains
71
+ else:
72
+ p["wb_channel"] = None
73
+ return p
74
+
75
+
76
+ def _conv(x, kernel):
77
+ n, c, h, w = x.shape
78
+ k = kernel.to(x)[None, None].repeat(c, 1, 1, 1)
79
+ pad = kernel.shape[-1] // 2
80
+ return F.conv2d(F.pad(x, (pad, pad, pad, pad), mode="reflect"), k, groups=c)
81
+
82
+
83
+ def motion_kernel(length, angle_deg):
84
+ size = max(3, int(round(length)) | 1)
85
+ k = torch.zeros(size, size)
86
+ c = size // 2
87
+ th = math.radians(angle_deg)
88
+ for s in np.linspace(-(length - 1) / 2, (length - 1) / 2, 8 * size):
89
+ k[int(round(c - s * math.sin(th))), int(round(c + s * math.cos(th)))] = 1.0
90
+ return k / k.sum()
91
+
92
+
93
+ def disk_kernel(r):
94
+ size = 2 * int(math.ceil(r)) + 1
95
+ c = size // 2
96
+ yy, xx = torch.meshgrid(torch.arange(size), torch.arange(size), indexing="ij")
97
+ k = (((yy - c) ** 2 + (xx - c) ** 2) <= (r + 0.5) ** 2).float()
98
+ return k / k.sum()
99
+
100
+
101
+ def apply(x, p, obs_mask=None, noise_seed=0):
102
+ t, v, c, h, w = x.shape
103
+ x = x.clone().clamp(0, 1)
104
+ for vi, (angle, scale) in enumerate(p["geometry"]):
105
+ x[:, vi] = TF.affine(x[:, vi], angle=angle, translate=[0, 0], scale=scale, shear=[0, 0],
106
+ interpolation=InterpolationMode.BILINEAR)
107
+ if p["light_shadow"] is not None:
108
+ yy, xx = torch.meshgrid(torch.linspace(0, 1, h), torch.linspace(0, 1, w), indexing="ij")
109
+ for vi, blobs in enumerate(p["light_shadow"]):
110
+ m = torch.ones(h, w)
111
+ for cx, cy, s, g in blobs:
112
+ m = m + (g - 1.0) * torch.exp(-((xx - cx) ** 2 + (yy - cy) ** 2) / (2 * s * s))
113
+ x[:, vi] = (x[:, vi] * m.clamp(min=0).to(x)).clamp(0, 1)
114
+ if p["wb_channel"] is not None:
115
+ for vi, g in enumerate(p["wb_channel"]):
116
+ x[:, vi] = (x[:, vi] * torch.tensor(g, dtype=x.dtype).view(1, 3, 1, 1)).clamp(0, 1)
117
+ y = x.reshape(t * v, c, h, w)
118
+ y = (y * float(2.0 ** p["exposure"])).clamp(0, 1)
119
+ d = p["temperature"]
120
+ y = (y * y.new_tensor([1.0 + d, 1.0, 1.0 - d]).view(1, 3, 1, 1)).clamp(0, 1)
121
+ y = ((y - 0.5) * p["contrast"] + 0.5).clamp(0, 1)
122
+ y = TF.adjust_saturation(y, p["saturation"])
123
+ y = TF.adjust_hue(y, p["hue"])
124
+ y = y.clamp(0, 1).pow(p["gamma"])
125
+ x = y.reshape(t, v, c, h, w)
126
+ mask = torch.ones(t, dtype=torch.bool) if obs_mask is None else torch.as_tensor(obs_mask, dtype=torch.bool)
127
+ if bool(mask.any()):
128
+ o = x[mask].reshape(-1, c, h, w)
129
+ if p["gaussian_blur"] > 0:
130
+ o = TF.gaussian_blur(o, [3, 3], [p["gaussian_blur"]] * 2)
131
+ if p["motion_blur"] is not None:
132
+ o = _conv(o, motion_kernel(*p["motion_blur"]))
133
+ if p["defocus"] > 0:
134
+ o = _conv(o, disk_kernel(p["defocus"]))
135
+ if p["sharpen"] > 0:
136
+ o = o + p["sharpen"] * (o - TF.gaussian_blur(o, [5, 5], [1.0, 1.0]))
137
+ o = o.clamp(0, 1)
138
+ if p["noise"] > 0:
139
+ gen = torch.Generator().manual_seed(int(noise_seed))
140
+ o = (o + torch.randn(o.shape, generator=gen, dtype=o.dtype) * p["noise"]).clamp(0, 1)
141
+ if p["jpeg"] is not None:
142
+ for i in range(o.shape[0]):
143
+ buf = io.BytesIO()
144
+ TF.to_pil_image(o[i]).save(buf, format="JPEG", quality=int(p["jpeg"]))
145
+ buf.seek(0)
146
+ with Image.open(buf) as im:
147
+ o[i] = TF.pil_to_tensor(im.convert("RGB")).to(o.dtype).div(255.0)
148
+ x[mask] = o.reshape(-1, v, c, h, w)
149
+ return x.clamp(0, 1)
150
+
151
+
152
+ def augment(x, rng, obs_mask=None):
153
+ p = sample(rng, x.shape[1], x.shape[-2])
154
+ return apply(x, p, obs_mask, noise_seed=int(rng.integers(0, 2 ** 62))), p
155
+
156
+
157
+ def perturb_joints(q, rng):
158
+ q = np.asarray(q, dtype=np.float64)
159
+ d = q.shape[-1]
160
+ bias = rng.uniform(-math.radians(2.0), math.radians(2.0), size=d)
161
+ sigma = math.radians(0.25)
162
+ noise = np.clip(rng.normal(0.0, sigma, size=q.shape), -3 * sigma, 3 * sigma)
163
+ return (q + bias + noise).astype(np.float32)
a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/hb_inloop_eval.py ADDED
@@ -0,0 +1,347 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """GR00T in-training open-loop action MSE (2026-11-15 user: every 250 training steps).
2
+
3
+ Imported by hb_groot/hb_launch.py (hb_launch_v.py runs hb_launch.py). It wraps Gr00tTrainer.__init__ so that every
4
+ trainer gets a TrainerCallback. At global_step % HB_INLOOP_EVERY == 0 (default 250; 0 = off) every rank evaluates its
5
+ shard of the benchmark eval cache (eval-action-mse/cache) with the weights being trained right now, exactly as deployed:
6
+ * Gr00tPolicy pipeline: eval processor from <output_dir>/processor (the files every checkpoint ships),
7
+ parse_observation_gr00t inputs, bf16 inputs, model.get_action under inference_mode, processor.decode_action;
8
+ * same protocol and metrics as the post-hoc eval (hb_eval_common): queries every 16 frames, first 16 predicted steps
9
+ vs the recorded actions, nMSE with the training-frame std, per group / per step;
10
+ * split: "heldout" when the task has recorded episodes that are never trained on (B6, trolley, B7), else "train10".
11
+ Training is not changed: module train/eval modes are restored module by module, torch CPU + CUDA (this device), numpy
12
+ and python RNG states are restored after each eval; the flow-matching noise uses a fixed seed (rank, batch) so every
13
+ eval sees the same noise (smoother curve). ZeRO-2 keeps the full bf16 weights on every rank, so ranks run independently;
14
+ partial sums are all-reduced once. Results: <run>/logs/inloop-mse.jsonl (rank 0), one json line per eval, and an
15
+ "HB_INLOOP" log line. Failures never stop training: an exception on any rank disables further evals on all ranks
16
+ (error flag in the same all-reduce, so no rank waits on another).
17
+ Env: HB_INLOOP_EVERY (250), HB_INLOOP_BATCH (8 queries per forward), HB_INLOOP_MAX_QUERIES (0 = all; tests only),
18
+ HB_INLOOP_SPLIT (override), HB_INLOOP_PROCESSOR_DIR (tests only).
19
+ """
20
+ import json
21
+ import os
22
+ import random
23
+ import sys
24
+ import time
25
+ import traceback
26
+ from contextlib import nullcontext
27
+ from pathlib import Path
28
+
29
+ import numpy as np
30
+ import torch
31
+
32
+ EVERY = int(os.environ.get("HB_INLOOP_EVERY", "250"))
33
+ BATCH = int(os.environ.get("HB_INLOOP_BATCH", "8"))
34
+ MAX_Q = int(os.environ.get("HB_INLOOP_MAX_QUERIES", "0"))
35
+ SEED = 1234
36
+ G = Path(__file__).resolve().parent.parent
37
+ EVAL_DIR = G / "eval-action-mse"
38
+ for _p in (str(EVAL_DIR), str(EVAL_DIR / "adapters")):
39
+ if _p not in sys.path:
40
+ sys.path.insert(0, _p)
41
+ import hb_eval_common as E # noqa: E402
42
+ import ad_groot as AG # noqa: E402 (canonical state -> GR00T state groups, same mapping as the post-hoc eval)
43
+
44
+ RANK = int(os.environ.get("RANK", "0"))
45
+ WORLD = int(os.environ.get("WORLD_SIZE", "1"))
46
+
47
+
48
+ def say(*a):
49
+ if RANK == 0:
50
+ print("HB_INLOOP", *a, flush=True)
51
+
52
+
53
+ def _argv(name):
54
+ a = sys.argv
55
+ for i, x in enumerate(a):
56
+ if x == name and i + 1 < len(a):
57
+ return a[i + 1]
58
+ if x.startswith(name + "="):
59
+ return x.split("=", 1)[1]
60
+ return None
61
+
62
+
63
+ def resolve(dataset_path):
64
+ """groot-data/<view> -> (benchmark task, split, view dir) or raises with the reason"""
65
+ view = Path(dataset_path)
66
+ rec = json.loads((view / "hb-dataset-receipt.json").read_text())
67
+ task = E.task_for_source(rec["source"])
68
+ if task is None:
69
+ raise LookupError(f"no benchmark task for source {rec['source']}")
70
+ split = os.environ.get("HB_INLOOP_SPLIT")
71
+ if not split:
72
+ split = "heldout" if (E.split_dir(task, "heldout") / "meta.json").exists() else "train10"
73
+ if not (E.split_dir(task, split) / "meta.json").exists():
74
+ raise FileNotFoundError(f"eval cache {task}/{split} missing (build_cache.py {task} {split})")
75
+ return task, split, view
76
+
77
+
78
+ class InLoopMSE:
79
+ """transformers TrainerCallback (duck-typed: the callback handler only calls the hooks that exist)"""
80
+
81
+ def __init__(self, task, split, view, processor_dir=None):
82
+ self.task, self.split, self.view = task, split, view
83
+ self.processor_dir = processor_dir
84
+ self.disabled = False
85
+ self.ready = False
86
+ self.results = []
87
+
88
+ # ---------------------------------------------------------------- setup (first eval; needs the live model)
89
+ def _setup(self, model, output_dir):
90
+ from transformers import AutoProcessor
91
+ from gr00t.data.embodiment_tags import EmbodimentTag
92
+ from gr00t.policy.gr00t_policy import Gr00tPolicy
93
+ from gr00t.policy.policy import BasePolicy
94
+
95
+ meta = E.load_meta(self.task, self.split)
96
+ self.meta = meta
97
+ self.lay = E.LAYOUTS[meta["layout"]]
98
+ self.std = E.norm_std(E.load_stats(self.task))
99
+ # the Gr00tPolicy of a checkpoint, built around the live model instead of loading weights from disk
100
+ pol = Gr00tPolicy.__new__(Gr00tPolicy)
101
+ BasePolicy.__init__(pol, strict=True)
102
+ pdir = Path(self.processor_dir or os.environ.get("HB_INLOOP_PROCESSOR_DIR") or Path(output_dir) / "processor")
103
+ proc = AutoProcessor.from_pretrained(str(pdir))
104
+ proc.eval()
105
+ proc.validate_strict_action_padding_mask(model.config)
106
+ tag = EmbodimentTag.NEW_EMBODIMENT
107
+ pol.model, pol.processor, pol.embodiment_tag = model, proc, tag
108
+ pol.modality_configs = {k: v for k, v in proc.get_modality_configs()[tag.value].items() if k != "rl_info"}
109
+ pol.collate_fn = proc.collator
110
+ pol.language_key = pol.modality_configs["language"].modality_keys[0]
111
+ self.policy = pol
112
+ self.mod = pol.modality_configs
113
+ self.state_map = AG.H2_STATE if meta["layout"].startswith("h2") else AG.G1_STATE
114
+ self.video_keys = list(self.mod["video"].modality_keys)
115
+ self.state_keys = list(self.mod["state"].modality_keys)
116
+ self.action_keys = list(self.mod["action"].modality_keys)
117
+ missing = [k for k in self.state_keys if k not in self.state_map] + \
118
+ [k for k in self.video_keys if k not in meta["cams"]]
119
+ if missing:
120
+ raise KeyError(f"model expects {missing}, not in the benchmark state / cache cameras {meta['cams']}")
121
+ tj = self.view / "meta/tasks.jsonl"
122
+ self.prompts = {str(r["task_index"]): r["task"] for r in E.jl(tj)} if tj.exists() else {}
123
+ # all queries of the split in a fixed order; this rank keeps every WORLD-th one (images only for those)
124
+ allq = []
125
+ for ep in meta["episodes"]:
126
+ with np.load(E.split_dir(self.task, self.split) / f"ep_{ep:06d}.npz") as z:
127
+ allq += [(ep, qi, int(t)) for qi, t in enumerate(z["query_t"])]
128
+ if MAX_Q:
129
+ allq = allq[:MAX_Q]
130
+ self.n_total = len(allq)
131
+ mine = [q for i, q in enumerate(allq) if i % WORLD == RANK]
132
+ self.eps = list(meta["episodes"])
133
+ self.q = [] # (ep, t, T, state[t], prompt, {cam: img}, gt(16, Da), valid(16))
134
+ by_ep = {}
135
+ for ep, qi, t in mine:
136
+ by_ep.setdefault(ep, []).append((qi, t))
137
+ for ep, lst in by_ep.items():
138
+ with np.load(E.split_dir(self.task, self.split) / f"ep_{ep:06d}.npz") as z:
139
+ state, action = z["state"], z["action"]
140
+ prompt = meta["prompts"][str(int(z["task_index"]))]
141
+ imgs = {c: z[f"img_{c}"] for c in self.video_keys}
142
+ for qi, t in lst:
143
+ gt, valid = E.gt_chunk(action, t)
144
+ self.q.append((ep, t, len(action), state[t].copy(), prompt,
145
+ {c: imgs[c][qi].copy() for c in self.video_keys}, gt, valid))
146
+ self.ready = True
147
+ say(f"ready: task {self.task} split {self.split} queries {self.n_total} (rank {RANK}: {len(self.q)}), "
148
+ f"every {EVERY} steps, batch {BATCH}, cams {self.video_keys}, processor {pdir}")
149
+
150
+ # ---------------------------------------------------------------- one forward on a batch of queries
151
+ def _obs(self, q):
152
+ from gr00t.data.utils import parse_observation_gr00t
153
+ ep, t, T, st, prompt, imgs, _, _ = q
154
+ o = {}
155
+ for k in self.state_keys:
156
+ a, b = self.state_map[k]
157
+ o[f"state.{k}"] = np.asarray(st[a:b], dtype=np.float32)[None]
158
+ for k in self.video_keys:
159
+ o[f"video.{k}"] = imgs[k][None]
160
+ o[self.language_key()] = self.prompts.get("0", prompt) if len(self.prompts) == 1 else prompt
161
+ return parse_observation_gr00t(o, self.mod)
162
+
163
+ def language_key(self):
164
+ return self.policy.language_key
165
+
166
+ @staticmethod
167
+ def _merge(obs_list):
168
+ out = {}
169
+ for m in obs_list[0]:
170
+ out[m] = {}
171
+ for k, v in obs_list[0][m].items():
172
+ vals = [o[m][k] for o in obs_list]
173
+ out[m][k] = np.concatenate(vals, axis=0) if isinstance(v, np.ndarray) else sum(vals, [])
174
+ return out
175
+
176
+ def _predict(self, batch, model):
177
+ obs = self._merge([self._obs(q) for q in batch])
178
+ dtype = next(model.parameters()).dtype
179
+ ctx = torch.autocast("cuda", dtype=torch.bfloat16) if (dtype != torch.bfloat16 and torch.cuda.is_available()) \
180
+ else nullcontext()
181
+ with ctx:
182
+ act, _ = self.policy.get_action(obs)
183
+ return np.concatenate([np.asarray(act[k], dtype=np.float64) for k in self.action_keys], axis=-1)
184
+
185
+ # ---------------------------------------------------------------- the eval
186
+ def evaluate(self, step, model, output_dir):
187
+ t0 = time.time()
188
+ D = self.lay["action_dim"] if self.ready else 0
189
+ err, msg = 0.0, ""
190
+ acc = None
191
+ nq = 0
192
+ infer = 0.0
193
+ modes = [(m, m.training) for m in model.modules()]
194
+ dev = torch.cuda.current_device() if torch.cuda.is_available() else None
195
+ np_state, py_state = np.random.get_state(), random.getstate()
196
+ try:
197
+ with torch.random.fork_rng(devices=[dev] if dev is not None else []):
198
+ if not self.ready:
199
+ self._setup(model, output_dir)
200
+ D = self.lay["action_dim"]
201
+ model.eval()
202
+ acc = E.Accumulator(self.std, self.lay["groups"])
203
+ for bi in range(0, len(self.q), BATCH):
204
+ batch = self.q[bi:bi + BATCH]
205
+ torch.default_generator.manual_seed(SEED + 1000 * RANK + bi)
206
+ if dev is not None:
207
+ torch.cuda.manual_seed(SEED + 1000 * RANK + bi)
208
+ s = time.time()
209
+ pred = self._predict(batch, model)
210
+ infer += time.time() - s
211
+ if pred.shape[0] != len(batch) or pred.shape[1] < E.HORIZON or pred.shape[2] != D:
212
+ raise ValueError(f"prediction {pred.shape}, need ({len(batch)}, >={E.HORIZON}, {D})")
213
+ for q, p in zip(batch, pred):
214
+ acc.add(q[0], p[:E.HORIZON], q[6], q[7])
215
+ nq += 1
216
+ except Exception as e: # noqa: BLE001 - never stop training
217
+ err, msg = 1.0, f"{type(e).__name__}: {e}"
218
+ print(f"HB_INLOOP rank {RANK} step {step} FAILED: {msg}\n{traceback.format_exc()}", file=sys.stderr,
219
+ flush=True)
220
+ finally:
221
+ for m, tr in modes:
222
+ m.training = tr
223
+ np.random.set_state(np_state)
224
+ random.setstate(py_state)
225
+ # ---- one all-reduce of all partial sums (+ error flag); D / eps are identical on every rank once ready
226
+ H = E.HORIZON
227
+ eps = self.eps if self.ready else []
228
+ nE = len(eps)
229
+ vec = np.zeros(3 * D + 1 + 2 * H + 2 * nE + 3)
230
+ if acc is not None and not err:
231
+ o = 0
232
+ vec[o:o + D] = acc.sse_raw; o += D
233
+ vec[o:o + D] = acc.sae_raw; o += D
234
+ vec[o:o + D] = acc.sse_n; o += D
235
+ vec[o] = acc.rows; o += 1
236
+ vec[o:o + H] = acc.step_sse_n; o += H
237
+ vec[o:o + H] = acc.step_rows; o += H
238
+ for j, ep in enumerate(eps):
239
+ s, c = acc.per_episode.get(int(ep), (0.0, 0))
240
+ vec[o + j], vec[o + nE + j] = s, c
241
+ o += 2 * nE
242
+ vec[o], vec[o + 1] = nq, infer
243
+ vec[-1] = err
244
+ tot = self._allreduce(vec)
245
+ if tot is None: # ranks disagree on the vector length (setup failed somewhere) -> stop for good
246
+ self.disabled = True
247
+ self._write(output_dir, {"step": step, "error": "length mismatch across ranks / collective failed"})
248
+ return
249
+ if tot[-1] > 0 or not self.ready:
250
+ self.disabled = True
251
+ self._write(output_dir, {"step": step, "error": msg or "failed on another rank", "ranks_failed": tot[-1]})
252
+ say(f"step {step}: eval failed on {int(tot[-1])} rank(s) -> in-training eval disabled ({msg})")
253
+ return
254
+ o = 0
255
+ acc = E.Accumulator(self.std, self.lay["groups"])
256
+ acc.sse_raw = tot[o:o + D]; o += D
257
+ acc.sae_raw = tot[o:o + D]; o += D
258
+ acc.sse_n = tot[o:o + D]; o += D
259
+ acc.rows = int(round(tot[o])); o += 1
260
+ acc.step_sse_n = tot[o:o + H]; o += H
261
+ acc.step_rows = tot[o:o + H]; o += H
262
+ acc.per_episode = {int(ep): [float(tot[o + j]), int(round(tot[o + nE + j]))]
263
+ for j, ep in enumerate(eps) if tot[o + nE + j] > 0}
264
+ o += 2 * nE
265
+ n_q, infer_sum = int(round(tot[o])), float(tot[o + 1])
266
+ res = acc.result()
267
+ res.pop("nmse_by_dim", None)
268
+ rec = {"step": step, "utc": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()), "task": self.task,
269
+ "split": self.split, "queries": n_q, "ranks": WORLD, "wall_s": round(time.time() - t0, 1),
270
+ "infer_s_per_query": round(infer_sum / max(n_q, 1), 4),
271
+ "denoising_steps": int(getattr(model.action_head, "num_inference_timesteps", -1))
272
+ if hasattr(model, "action_head") else None, "seed": SEED, **res}
273
+ self.results.append(rec)
274
+ self._write(output_dir, rec)
275
+ g = " ".join(f"{k} {v['nmse']:.4f}" for k, v in res["groups"].items())
276
+ say(f"step {step}: {self.task}/{self.split} nMSE {res['nmse']:.4f} ({g}) on {n_q} queries, "
277
+ f"{rec['wall_s']} s")
278
+
279
+ def _allreduce(self, vec):
280
+ import torch.distributed as dist
281
+ if not (dist.is_available() and dist.is_initialized()) or WORLD == 1:
282
+ return vec
283
+ try:
284
+ dev = "cuda" if dist.get_backend() == "nccl" else "cpu"
285
+ n = torch.tensor([float(len(vec))], dtype=torch.float64, device=dev)
286
+ mx = n.clone()
287
+ dist.all_reduce(mx, op=dist.ReduceOp.MAX)
288
+ mn = n.clone()
289
+ dist.all_reduce(mn, op=dist.ReduceOp.MIN)
290
+ if mx.item() != mn.item():
291
+ return None
292
+ t = torch.from_numpy(vec).to(device=dev, dtype=torch.float64)
293
+ dist.all_reduce(t)
294
+ return t.cpu().numpy()
295
+ except Exception as e: # noqa: BLE001
296
+ print(f"HB_INLOOP rank {RANK} all_reduce failed: {e!r}", file=sys.stderr, flush=True)
297
+ return None
298
+
299
+ def _write(self, output_dir, rec):
300
+ if RANK != 0:
301
+ return
302
+ try:
303
+ out = Path(output_dir).parent / "logs" / "inloop-mse.jsonl"
304
+ out.parent.mkdir(parents=True, exist_ok=True)
305
+ with open(out, "a") as f:
306
+ f.write(json.dumps(rec) + "\n")
307
+ except Exception as e: # noqa: BLE001
308
+ print(f"HB_INLOOP write failed: {e!r}", flush=True)
309
+
310
+ # ---------------------------------------------------------------- TrainerCallback hooks
311
+ def on_step_end(self, args, state, control, model=None, **kw):
312
+ step = int(state.global_step)
313
+ if self.disabled or EVERY <= 0 or step <= 0 or step % EVERY or model is None:
314
+ return control
315
+ self.evaluate(step, model, args.output_dir)
316
+ return control
317
+
318
+
319
+ def _install():
320
+ if EVERY <= 0:
321
+ say("disabled (HB_INLOOP_EVERY=0)")
322
+ return
323
+ ds = _argv("--dataset_path") or _argv("--dataset-path")
324
+ if not ds:
325
+ say("disabled: no --dataset_path in argv")
326
+ return
327
+ try:
328
+ task, split, view = resolve(ds)
329
+ except Exception as e: # noqa: BLE001
330
+ say(f"disabled: {e}")
331
+ return
332
+ from transformers import TrainerCallback
333
+ from gr00t.experiment.trainer import Gr00tTrainer
334
+
335
+ cb_cls = type("HBInLoopMSE", (InLoopMSE, TrainerCallback), {})
336
+ _orig_init = Gr00tTrainer.__init__
337
+
338
+ def __init__(self, *a, **k):
339
+ _orig_init(self, *a, **k)
340
+ self.add_callback(cb_cls(task, split, view))
341
+
342
+ Gr00tTrainer.__init__ = __init__
343
+ say(f"installed: task {task} split {split} every {EVERY} steps (view {view.name})")
344
+
345
+
346
+ if os.environ.get("HB_INLOOP_NO_INSTALL") != "1":
347
+ _install()
a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/hb_launch.py ADDED
@@ -0,0 +1,100 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """GR00T N1.7 fine-tuning entry with Humanoid Benchmark augmentation v2.
2
+
3
+ Patches Gr00tN1d7Processor in training mode only:
4
+ - images: native random crop / colour jitter / rotation are bypassed (the deterministic eval resize+center-crop is
5
+ used), then hb_aug_v2 is applied jointly to all cameras of the sample;
6
+ - proprio: the body-joint state groups (HB_BODY_STATE_KEYS) get the v2 bias + noise before normalization.
7
+ Then runs the pinned upstream gr00t/experiment/launch_finetune.py unchanged.
8
+ """
9
+ import json
10
+ import os
11
+ import runpy
12
+ import sys
13
+ from pathlib import Path
14
+
15
+ import numpy as np
16
+ import torch
17
+
18
+ HERE = Path(__file__).resolve().parent
19
+ sys.path.insert(0, str(HERE))
20
+ import hb_offline_fix # noqa: E402,F401 (tokenizer load under HF_HUB_OFFLINE=1, see module docstring)
21
+ import hb_aug_v2 as A # noqa: E402
22
+ from gr00t.model.gr00t_n1d7 import processing_gr00t_n1d7 as P # noqa: E402
23
+
24
+ Proc = P.Gr00tN1d7Processor
25
+ BODY_KEYS = [k for k in os.environ.get("HB_BODY_STATE_KEYS", "left_leg,right_leg,waist,left_arm,right_arm").split(",") if k]
26
+ LOG = os.environ.get("HB_AUG_LOG")
27
+ _state = {"rng": None, "count": 0}
28
+
29
+
30
+ def _rng():
31
+ if _state["rng"] is None:
32
+ _state["rng"] = np.random.default_rng([os.getpid(), int.from_bytes(os.urandom(8), "little")])
33
+ return _state["rng"]
34
+
35
+
36
+ _orig_get, _orig_apply, _orig_call = Proc._get_vlm_inputs, Proc._apply_vlm_processing, Proc.__call__
37
+
38
+
39
+ def _get_vlm_inputs(self, image_keys, images, masks, image_transform, language):
40
+ if getattr(self, "training", False):
41
+ self._hb_views = len(image_keys)
42
+ try:
43
+ return _orig_get(self, image_keys, images, masks, self.eval_image_transform, language)
44
+ finally:
45
+ self._hb_views = None
46
+ return _orig_get(self, image_keys, images, masks, image_transform, language)
47
+
48
+
49
+ def _apply_vlm_processing(self, images, language):
50
+ views = getattr(self, "_hb_views", None)
51
+ if views and isinstance(images, torch.Tensor):
52
+ shape = images.shape
53
+ x = images.reshape(-1, views, *shape[1:]).float().div(255.0)
54
+ x, p = A.augment(x, _rng())
55
+ images = (x * 255.0 + 0.5).clamp(0, 255).to(torch.uint8).reshape(shape)
56
+ _state["count"] += 1
57
+ if LOG and _state["count"] in (1, 50):
58
+ with open(f"{LOG}.{os.getpid()}", "a") as f:
59
+ f.write(json.dumps({"count": _state["count"], "views": views, "hw": list(shape[-2:]),
60
+ "boost": p["boost"], "version": A.VERSION}) + "\n")
61
+ return _orig_apply(self, images, language)
62
+
63
+
64
+ def _call(self, messages):
65
+ if getattr(self, "training", False) and BODY_KEYS:
66
+ states = messages[0]["content"].states
67
+ for key in BODY_KEYS:
68
+ if key in states:
69
+ states[key] = A.perturb_joints(states[key], _rng())
70
+ return _orig_call(self, messages)
71
+
72
+
73
+ Proc._get_vlm_inputs = _get_vlm_inputs
74
+ Proc._apply_vlm_processing = _apply_vlm_processing
75
+ Proc.__call__ = _call
76
+
77
+ # 2026-11-15 user: in-training open-loop action MSE every 250 steps (hb_inloop_eval.py adds a TrainerCallback to
78
+ # Gr00tTrainer; HB_INLOOP_EVERY=0 turns it off). It never blocks training: any setup problem only disables it.
79
+ if os.environ.get("HB_PATCH_ONLY") != "1":
80
+ try:
81
+ import hb_inloop_eval # noqa: E402,F401
82
+ except Exception as _e: # noqa: BLE001
83
+ print(f"HB_INLOOP disabled: {_e!r}", flush=True)
84
+
85
+ if os.environ.get("HB_PATCH_ONLY") != "1":
86
+ entry = Path(os.environ["GROOT_SRC"]) / "gr00t/experiment/launch_finetune.py"
87
+ sys.argv = [str(entry)] + sys.argv[1:]
88
+ try:
89
+ runpy.run_path(str(entry), run_name="__main__")
90
+ finally: # batch probe: report this rank's peak reserved memory (groot_chain.py compares it with MAX_FRAC)
91
+ _peak = os.environ.get("HB_PEAK_FILE")
92
+ if _peak and torch.cuda.is_available():
93
+ try:
94
+ _dev = torch.cuda.current_device()
95
+ with open(f"{_peak}-rank{os.environ.get('RANK', '0')}.json", "w") as _f:
96
+ json.dump({"rank": int(os.environ.get("RANK", "0")),
97
+ "peak_reserved_gb": torch.cuda.max_memory_reserved(_dev) / 1e9,
98
+ "total_gb": torch.cuda.get_device_properties(_dev).total_memory / 1e9}, _f)
99
+ except Exception: # noqa: BLE001 - never mask the training result
100
+ pass
a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/hb_launch_v.py ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """2026-11-15 user: GR00T variant launcher = hb_launch.py (HB aug v2 + A20 proprio + processor patch, then the pinned
2
+ upstream launch_finetune.py) plus, from environment variables, before the run starts:
3
+ HB_GROOT_MODEL_OVERRIDES JSON {"attn_dropout": 0.0, "diffusion_model_cfg.dropout": 0.0, "num_inference_timesteps": 16,
4
+ ...}: set on config.model (dotted keys go into nested dicts / attributes)
5
+ HB_GROOT_LR_GROUPS JSON {"language": 1e-6, "vision": 1e-5, "projector": 1e-4, "rest": 1e-4}: layer-wise learning
6
+ rates (Psi-Zero style) through Gr00tTrainer.create_optimizer:
7
+ language = backbone ... language_model.* and lm_head
8
+ vision = backbone ... visual.* except the vision->language mergers
9
+ projector = visual.merger.* and visual.deepstack_merger_list.*
10
+ rest = action head (DiT, VL self-attention, VLLN, encoders / decoders)
11
+ weight decay as HF Trainer (no decay on biases / norms); cosine + warmup scale every group.
12
+ Prints HB_LAUNCH_V lines (overrides applied, parameter count and lr per group) on rank 0 for verification."""
13
+ import json
14
+ import os
15
+ import runpy
16
+ import sys
17
+ from pathlib import Path
18
+
19
+ H = Path(__file__).resolve().parent
20
+ sys.path.insert(0, str(H))
21
+ RANK0 = os.environ.get("RANK", "0") == "0"
22
+
23
+
24
+ def say(*a):
25
+ if RANK0:
26
+ print("HB_LAUNCH_V", *a, flush=True)
27
+
28
+
29
+ OV = json.loads(os.environ.get("HB_GROOT_MODEL_OVERRIDES", "{}") or "{}")
30
+ LRG = json.loads(os.environ.get("HB_GROOT_LR_GROUPS", "{}") or "{}")
31
+
32
+ import gr00t.experiment.experiment as E # noqa: E402
33
+
34
+ _run = E.run
35
+
36
+
37
+ def _set(obj, dotted, value):
38
+ parts = dotted.split(".")
39
+ for p in parts[:-1]:
40
+ obj = obj[p] if isinstance(obj, dict) else getattr(obj, p)
41
+ last = parts[-1]
42
+ if isinstance(obj, dict):
43
+ if last not in obj:
44
+ raise KeyError(f"model override {dotted}: no key {last}")
45
+ old = obj[last]
46
+ obj[last] = value
47
+ else:
48
+ if not hasattr(obj, last):
49
+ raise AttributeError(f"model override {dotted}: no attribute {last}")
50
+ old = getattr(obj, last)
51
+ setattr(obj, last, value)
52
+ say(f"override model.{dotted}: {old!r} -> {value!r}")
53
+
54
+
55
+ def run(config):
56
+ for k, v in OV.items():
57
+ _set(config.model, k, v)
58
+ return _run(config)
59
+
60
+
61
+ E.run = run
62
+
63
+ if LRG:
64
+ from transformers import Trainer
65
+ from gr00t.experiment.trainer import Gr00tTrainer
66
+
67
+ def group_of(name):
68
+ if ".language_model." in name or "lm_head" in name:
69
+ return "language"
70
+ if ".visual." in name:
71
+ return "projector" if (".visual.merger." in name or ".visual.deepstack_merger_list." in name) else "vision"
72
+ return "rest"
73
+
74
+ def create_optimizer(self):
75
+ if self.optimizer is not None:
76
+ return self.optimizer
77
+ model = self.model
78
+ decay = set(self.get_decay_parameter_names(model))
79
+ groups, stats = {}, {}
80
+ for n, p in model.named_parameters():
81
+ if not p.requires_grad:
82
+ continue
83
+ g = group_of(n)
84
+ lr = float(LRG.get(g, LRG.get("rest", self.args.learning_rate)))
85
+ wd = self.args.weight_decay if n in decay else 0.0
86
+ groups.setdefault((g, lr, wd), []).append(p)
87
+ s = stats.setdefault(g, [lr, 0, 0])
88
+ s[1] += 1
89
+ s[2] += p.numel()
90
+ opt_cls, opt_kwargs = Trainer.get_optimizer_cls_and_kwargs(self.args, model)
91
+ opt_kwargs.pop("lr", None)
92
+ self.optimizer = opt_cls([{"params": ps, "lr": lr, "weight_decay": wd} for (g, lr, wd), ps in groups.items()],
93
+ lr=self.args.learning_rate, **opt_kwargs)
94
+ for g, (lr, n_t, n_p) in sorted(stats.items()):
95
+ say(f"lr group {g}: lr {lr:g}, {n_t} tensors, {n_p / 1e6:.1f} M params")
96
+ return self.optimizer
97
+
98
+ Gr00tTrainer.create_optimizer = create_optimizer
99
+ say("layer-wise lr groups installed", json.dumps(LRG))
100
+
101
+ say("model overrides", json.dumps(OV))
102
+ runpy.run_path(str(H / "hb_launch.py"), run_name="__main__")
a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/hb_offline_fix.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Make tokenizer loading work with HF_HUB_OFFLINE=1 on transformers 4.57.x.
2
+
3
+ PreTrainedTokenizerBase._patch_mistral_regex calls huggingface_hub.model_info() for every hub-named tokenizer with a
4
+ vocab > 100k (to detect Mistral base models). With HF_HUB_OFFLINE=1 that raises OfflineModeIsEnabled even though all
5
+ files are in the cache. GR00T's Cosmos-Reason2-2B / Qwen3-VL tokenizer is not a Mistral tokenizer, so when offline the
6
+ check is skipped and the tokenizer is returned unchanged. Online behaviour is untouched.
7
+ """
8
+ import os
9
+
10
+ from transformers.tokenization_utils_base import PreTrainedTokenizerBase as _PTB
11
+
12
+ if not getattr(_PTB, "_hb_offline_fix", False):
13
+ _orig = _PTB._patch_mistral_regex.__func__
14
+
15
+ def _patch_mistral_regex(cls, tokenizer, *args, **kwargs):
16
+ try:
17
+ return _orig(cls, tokenizer, *args, **kwargs)
18
+ except Exception as error:
19
+ if os.environ.get("HF_HUB_OFFLINE") == "1" or type(error).__name__ == "OfflineModeIsEnabled":
20
+ return tokenizer
21
+ raise
22
+
23
+ _PTB._patch_mistral_regex = classmethod(_patch_mistral_regex)
24
+ _PTB._hb_offline_fix = True
a18-pour-water/b512-2cam-h32-cropped-v2/step-004000/code/hb_groot/prepare_groot_data.py ADDED
@@ -0,0 +1,191 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Build GR00T dataset views (symlinked episodes + rewritten meta + stats) for the G1 tasks.
2
+
3
+ Usage: <groot python> prepare_groot_data.py <task> [<task> ...] (tasks: b6 basket a3 a7)
4
+ B6 always uses the fixed first 100 episodes (0-99) of b6-v5t7-ready-start-20261109.
5
+ """
6
+ import hashlib
7
+ import json
8
+ import os
9
+ import subprocess
10
+ import sys
11
+ from pathlib import Path
12
+
13
+ H = Path(__file__).resolve().parent
14
+ G = H.parent
15
+ P = G.parent
16
+ OUT = G / "groot-data"
17
+ TASKS = {
18
+ "b6": (P / "b6-v5t7-ready-start-60ep-20261109/data/deliveries/g1/b6-v5t7-ready-start-20261109", 100),
19
+ "basket": (P / "basket-current-dynamics-100-v21-b0ff522/data", None), # HF basket-transfer-expert-v1 @current-dynamics-100-v21 b0ff522 (old ed77719 retired)
20
+ "basketv2": (P / "basket-trimv2-c4df4d4/data", None), # HF a20-basket-transfer-pre-reach-trim-expert-v2 @c4df4d40 (user 2026-11-14)
21
+ "trolley": (P / "trolley-push-expert-v1-20261110/data/trolley_push", 100), # HF trolley-push-expert-v1 @ef9024a: fixed first 100 of 103
22
+ "b7": (P / "b7-black-clothes-hamper-v1-1bb002a/data", 100), # HF b7-black-clothes-hamper-expert-v1 @1bb002a, first 100 of 107
23
+ "a6": (P / "a6-door-pull-v1-cf3ee9b/data/door_pull", None), # HF door-pull-expert-v1 @cf3ee9b (A6)
24
+ "b7green": (P / "b7-green-clothes-hamper-v1-ada288c/data", None), # HF b7-green-clothes-hamper-expert-v1 @ada288c, 100 episodes (user 2026-11-13)
25
+ "a7v2": (P / "a7-cabinet-close-v2-aecf070/data", None), # HF a7-cabinet-close-expert-v2 @aecf070, 100 episodes (user 2026-11-14)
26
+ "b8": (P / "b8-table-wipe-v1-3b2e618/data", None), # HF b8-table-wipe-expert-v1 @3b2e618, 100 episodes (user 2026-11-14)
27
+ "b1": (P / "b1-riceball-microwave-v1-36da298/data", None), # HF b1-riceball-microwave-expert-v1 @36da298, all 115 episodes (user 2026-11-14)
28
+ "a10v3": (P / "a10-trolley-push-v3-6e410c5/data", None), # HF a10-trolley-push-expert-v3 @6e410c5, all 148 episodes (user 2026-11-15)
29
+ "a3v3": (P / "a3-fridge-open-v3-24882d5/data", None), # HF a3-fridge-open-expert-v3 @24882d5, all 125 episodes, measured hands (user 2026-11-15)
30
+ "a7v3": (P / "a7-cabinet-close-v3-c2e15b0/data", None), # HF a7-cabinet-close-expert-v3 @c2e15b0, all 100 episodes, measured hands (user 2026-11-15)
31
+ "b6v6": (P / "b6-pedal-bin-v6-ad2f5bf/data", None), # HF b6-pedal-bin @ad2f5bf0 (v6 at the repo root), all 542 episodes (user 2026-11-15)
32
+ "b5": (P / "b5-chair-push-v1-b85b7e9/data", None), # HF b5-chair-push-expert-v1 @b85b7e9, all 100 episodes (user 2026-11-15)
33
+ "a3v4": (P / "a3-fridge-open-v4-6393d4a/data", None), # HF a3-fridge-open-expert-v4 @6393d4a, all 125 episodes (user 2026-11-15: v4 replaces v3)
34
+ "a7v4": (P / "a7-cabinet-close-v4-90cc2b5/data", None), # HF a7-cabinet-close-expert-v4 @90cc2b5, all 100 episodes (user 2026-11-15: v4 replaces v3)
35
+ "b7greenv2": (P / "b7-green-clothes-hamper-v2-c12757e/data", None), # HF b7-green-clothes-hamper @c12757e (v2 at the repo root), all 120 episodes (user 2026-11-14)
36
+ "a10v2": (P / "a10-trolley-push-v2-d3abb4c/data", None), # HF a10-trolley-push-expert-v2 @d3abb4c, all 150 episodes (user 2026-11-13)
37
+ "a1": (P / "a1-ground-object-pickup-v1-f47f07e/data", None), # HF a1-ground-object-pickup-expert-v1 @f47f07e (A1)
38
+ "c6t": (P / "c6-test8-06c231d/data", None), # c6 8-episode adaptation test only
39
+ "c6": (P / "c6-bathroom-door-fetch-v1-3afba69/data", None), # HF c6-bathroom-door-fetch-expert-v1 @3afba69 (C6)
40
+ "a3": (P / "a3-a7-20261109/a3-fridge-open-expert-v1/780fcf1e4bd7db947a8c2d8f7b92e61f24f71b45", None),
41
+ "a7": (P / "a3-a7-20261109/a7-cabinet-close-expert-v1/1935a98e8f55d4d1d2d1615ddb97c2679713595c", None),
42
+ "a18": (P / "a18-h2-dataset-audit-8e956a10/dataset", None),
43
+ "a17": (P / "a17-unscrew-bottle-v1-407e6e2/data", None), # HF a17-unscrew-bottle-expert-v1 @407e6e2
44
+ "a18c2": (P / "a18-h2-dataset-audit-8e956a10/dataset", None), # A18, 2 cameras head + chest (user 2026-11-14 18:2x)
45
+ "a17c2": (P / "a17-unscrew-bottle-v1-407e6e2/data", None), # A17, 2 cameras head + chest (user 2026-11-14 18:2x)
46
+ "a18cr2": (P / "a18-pour-water-cropped-v2-44cdd57/data", None), # HF a18-pour-water-expert-cropped-v2 @44cdd57, head + chest (user 2026-11-15)
47
+ "a17cr2": (P / "a17-unscrew-bottle-cropped-v2-3520b70/data", None), # HF a17-unscrew-bottle-expert-cropped-v2 @3520b70, head + chest (user 2026-11-15)
48
+ "a14": (P / "a14-unscrew-bulb-v1-b4390fc/data", None), # HF a14-unscrew-bulb-expert-v1 @b4390fc (H2, 3 cameras)
49
+ "a14v2": (P / "a14-unscrew-bulb-v2-8411a6d/data", None), # HF a14-unscrew-bulb-expert-v2 @8411a6da: + action.planner (H2 guide 2026-11-14)
50
+ }
51
+ CONFIG = {"a18": "h2w3_body31_hand44_h16.py", "a17": "h2w3_body31_hand44_h16.py", "a14": "h2w3_body31_hand44_h16.py",
52
+ "a18c2": "h2_body31_hand44_h16.py", "a17c2": "h2_body31_hand44_h16.py",
53
+ "a18cr2": "h2_body31_hand44_h16.py", "a17cr2": "h2_body31_hand44_h16.py",
54
+ "a14v2": "h2_body31_hand44_planner7_h16.py"} # 2026-11-14 H2 guide: action82 incl. the planner command
55
+ # model-input cameras per H2 task (user 2026-11-12 22:0x: EVERY H2 task uses head + chest + wrist; third_person never)
56
+ CAMS = {"a18": ("head", "chest", "wrist"), "a17": ("head", "chest", "wrist"), "a14": ("head", "chest", "wrist"),
57
+ "a18c2": ("head", "chest"), "a17c2": ("head", "chest"), "a18cr2": ("head", "chest"), "a17cr2": ("head", "chest"), "a14v2": ("head", "chest")} # 2026-11-14: 2-camera re-train (user)
58
+ PLANNER = {"a14v2"} # tasks whose action adds action.planner (7) after body_target + hands (H2 guide a14 2.2)
59
+ H2_GROUPS = {"left_leg": (0, 6), "right_leg": (6, 12), "waist": (12, 15), "left_arm": (15, 22), "right_arm": (22, 29),
60
+ "head": (29, 31), "left_hand": (31, 53), "right_hand": (53, 75)}
61
+
62
+
63
+ def h2_modality(info, cams=("head", "chest")):
64
+ f = info["features"]
65
+ names = f["observation.state"]["names"]
66
+ assert len(names) == 75 and names[:31] == f["action.body_target"]["names"] and names[31:] == f["action.hands"]["names"]
67
+ assert f["observation.projected_gravity"]["shape"] == [3]
68
+ state = {g: {"original_key": "observation.state", "start": a, "end": b} for g, (a, b) in H2_GROUPS.items()}
69
+ state["projected_gravity"] = {"original_key": "observation.projected_gravity", "start": 0, "end": 3}
70
+ return {"state": state,
71
+ "action": {"body_target": {"original_key": "action.body_target", "start": 0, "end": 31},
72
+ "left_hand_target": {"original_key": "action.hands", "start": 0, "end": 22},
73
+ "right_hand_target": {"original_key": "action.hands", "start": 22, "end": 44}},
74
+ "video": {c: {"original_key": f"observation.images.{c}"} for c in cams},
75
+ "annotation": {"human.task_description": {"original_key": "task_index"}}}
76
+ BODY = ["left_hip_pitch_joint", "left_hip_roll_joint", "left_hip_yaw_joint", "left_knee_joint", "left_ankle_pitch_joint",
77
+ "left_ankle_roll_joint", "right_hip_pitch_joint", "right_hip_roll_joint", "right_hip_yaw_joint",
78
+ "right_knee_joint", "right_ankle_pitch_joint", "right_ankle_roll_joint", "waist_yaw_joint", "waist_roll_joint",
79
+ "waist_pitch_joint", "left_shoulder_pitch_joint", "left_shoulder_roll_joint", "left_shoulder_yaw_joint",
80
+ "left_elbow_joint", "left_wrist_roll_joint", "left_wrist_pitch_joint", "left_wrist_yaw_joint",
81
+ "right_shoulder_pitch_joint", "right_shoulder_roll_joint", "right_shoulder_yaw_joint", "right_elbow_joint",
82
+ "right_wrist_roll_joint", "right_wrist_pitch_joint", "right_wrist_yaw_joint"]
83
+ GROUPS = {"left_leg": (0, 6), "right_leg": (6, 12), "waist": (12, 15), "left_arm": (15, 22), "right_arm": (22, 29)}
84
+
85
+
86
+ def jl(path):
87
+ return [json.loads(x) for x in Path(path).read_text().splitlines() if x.strip()]
88
+
89
+
90
+ def wl(path, rows):
91
+ Path(path).write_text("".join(json.dumps(r) + "\n" for r in rows))
92
+
93
+
94
+ def build(task):
95
+ src, limit = TASKS[task]
96
+ info = json.loads((src / "meta/info.json").read_text())
97
+ names = info["features"]["observation.state"]["names"]
98
+ if task in CONFIG:
99
+ idx = list(range(31))
100
+ else:
101
+ idx = [names.index(n) for n in BODY]
102
+ for g, (a, b) in GROUPS.items():
103
+ assert idx[a:b] == list(range(idx[a], idx[a] + (b - a))), (task, g, idx[a:b])
104
+ for key, dim in (("action.motion_token", 64), ("observation.inspire_left_hand_state", 6),
105
+ ("observation.inspire_right_hand_state", 6), ("observation.projected_gravity", 3),
106
+ ("observation.images.ego_view", None), ("observation.images.chest", None)):
107
+ assert key in info["features"], (task, key)
108
+ if dim:
109
+ assert info["features"][key]["shape"] == [dim], (task, key)
110
+ episodes = jl(src / "meta/episodes.jsonl")
111
+ stats = jl(src / "meta/episodes_stats.jsonl")
112
+ n = len(episodes) if limit is None else limit
113
+ episodes, stats = episodes[:n], stats[:n]
114
+ assert [e["episode_index"] for e in episodes] == list(range(n)) == [s["episode_index"] for s in stats]
115
+ frames = sum(int(e["length"]) for e in episodes)
116
+ dst = OUT / task
117
+ if dst.exists():
118
+ raise SystemExit(f"{dst} exists")
119
+ (dst / "meta").mkdir(parents=True)
120
+ vkeys = [k for k, f in info["features"].items() if f.get("dtype") == "video"]
121
+ for e in episodes:
122
+ i = e["episode_index"]
123
+ chunk = i // int(info["chunks_size"])
124
+ rel = info["data_path"].format(episode_chunk=chunk, episode_index=i)
125
+ (dst / rel).parent.mkdir(parents=True, exist_ok=True)
126
+ os.symlink(src / rel, dst / rel)
127
+ for k in vkeys:
128
+ relv = info["video_path"].format(episode_chunk=chunk, video_key=k, episode_index=i)
129
+ (dst / relv).parent.mkdir(parents=True, exist_ok=True)
130
+ os.symlink(src / relv, dst / relv)
131
+ if task in CONFIG:
132
+ # stats.py computes stats for every low-dim feature in info.json and cannot handle e.g. observation.tactile
133
+ # [10,24,24]; keep only what the H2 modality uses (parquet columns stay untouched behind the symlinks).
134
+ keep = {"observation.state", "observation.projected_gravity", "action.body_target", "action.hands",
135
+ "timestamp", "frame_index", "episode_index", "index", "task_index"}
136
+ keep |= {f"observation.images.{c}" for c in CAMS[task]}
137
+ keep |= {"action.planner"} if task in PLANNER else set()
138
+ info["features"] = {k: v for k, v in info["features"].items() if k in keep}
139
+ vkeys = [k for k in vkeys if k in keep]
140
+ stats = [{**s, "stats": {k: v for k, v in s["stats"].items() if k in keep}} if "stats" in s else s
141
+ for s in stats]
142
+ info.update(total_episodes=n, total_frames=frames, total_videos=n * len(vkeys),
143
+ total_chunks=(n - 1) // int(info["chunks_size"]) + 1, splits={"train": f"0:{n}"})
144
+ (dst / "meta/info.json").write_text(json.dumps(info, indent=2))
145
+ wl(dst / "meta/episodes.jsonl", episodes)
146
+ wl(dst / "meta/episodes_stats.jsonl", stats)
147
+ (dst / "meta/tasks.jsonl").write_text((src / "meta/tasks.jsonl").read_text())
148
+ if task in CONFIG:
149
+ modality = h2_modality(info, CAMS[task])
150
+ if task in PLANNER:
151
+ assert info["features"]["action.planner"]["shape"] == [7]
152
+ modality["action"]["planner"] = {"original_key": "action.planner", "start": 0, "end": 7}
153
+ else:
154
+ state = {g: {"original_key": "observation.state", "start": idx[a], "end": idx[a] + (b - a)}
155
+ for g, (a, b) in GROUPS.items()}
156
+ state.update(left_hand={"original_key": "observation.inspire_left_hand_state", "start": 0, "end": 6},
157
+ right_hand={"original_key": "observation.inspire_right_hand_state", "start": 0, "end": 6},
158
+ projected_gravity={"original_key": "observation.projected_gravity", "start": 0, "end": 3})
159
+ modality = {"state": state,
160
+ "action": {"motion_token": {"original_key": "action.motion_token", "start": 0, "end": 64},
161
+ "left_hand_joints": {"original_key": "observation.inspire_left_hand_state", "start": 0, "end": 6},
162
+ "right_hand_joints": {"original_key": "observation.inspire_right_hand_state", "start": 0, "end": 6}},
163
+ "video": {"ego_view": {"original_key": "observation.images.ego_view"},
164
+ "chest": {"original_key": "observation.images.chest"}},
165
+ "annotation": {"human.task_description": {"original_key": "task_index"}}}
166
+ (dst / "meta/modality.json").write_text(json.dumps(modality, indent=2))
167
+ py = sys.executable
168
+ src_g = G / "HumanWAM/third_party/Isaac-GR00T"
169
+ env = dict(os.environ, PYTHONPATH=f"{src_g}:{H}")
170
+ done = None
171
+ for tag in ("NEW_EMBODIMENT", "new_embodiment"):
172
+ r = subprocess.run([py, str(src_g / "gr00t/data/stats.py"), "--dataset-path", str(dst), "--embodiment-tag", tag,
173
+ "--modality-config-path", str(H / CONFIG.get(task, "g1_feedback76_h16.py"))], env=env, cwd=src_g,
174
+ capture_output=True, text=True)
175
+ if r.returncode == 0:
176
+ done = tag
177
+ break
178
+ last = r.stdout[-1500:] + r.stderr[-1500:]
179
+ if done is None:
180
+ raise SystemExit(f"stats failed for {task}:\n{last}")
181
+ assert (dst / "meta/stats.json").is_file()
182
+ receipt = dict(task=task, source=str(src), episodes=n, frames=frames, video_keys=vkeys,
183
+ state_dim_raw=len(names), body_index=idx, stats_tag=done,
184
+ modality_sha256=hashlib.sha256((dst / "meta/modality.json").read_bytes()).hexdigest(),
185
+ stats_sha256=hashlib.sha256((dst / "meta/stats.json").read_bytes()).hexdigest())
186
+ (dst / "hb-dataset-receipt.json").write_text(json.dumps(receipt, indent=2) + "\n")
187
+ print("DATASET_OK", json.dumps(receipt)[:400], flush=True)
188
+
189
+
190
+ for t in sys.argv[1:]:
191
+ build(t)