BIue-Jay commited on
Commit
3005355
·
verified ·
1 Parent(s): abefd49

a7: 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. a7-cabinet-close/b256/step-002000/code/hb_groot/cache_cosmos_reason2.py +60 -0
  2. a7-cabinet-close/b256/step-002000/code/hb_groot/check_reason2_offline.py +26 -0
  3. a7-cabinet-close/b256/step-002000/code/hb_groot/g1_feedback76_h16.py +27 -0
  4. a7-cabinet-close/b256/step-002000/code/hb_groot/groot_chain.py +139 -0
  5. a7-cabinet-close/b256/step-002000/code/hb_groot/h2_body31_hand44_h16.py +24 -0
  6. a7-cabinet-close/b256/step-002000/code/hb_groot/h2w3_body31_hand44_h16.py +24 -0
  7. a7-cabinet-close/b256/step-002000/code/hb_groot/hb_aug_v2.py +163 -0
  8. a7-cabinet-close/b256/step-002000/code/hb_groot/hb_launch.py +92 -0
  9. a7-cabinet-close/b256/step-002000/code/hb_groot/hb_offline_fix.py +24 -0
  10. a7-cabinet-close/b256/step-002000/code/hb_groot/prepare_groot_data.py +162 -0
  11. a7-cabinet-close/b256/step-002000/code/hb_groot/verify_groot_cpu.py +51 -0
  12. a7-cabinet-close/b256/step-002000/config.json +90 -0
  13. a7-cabinet-close/b256/step-002000/dataset_meta/hb-dataset-receipt.json +45 -0
  14. a7-cabinet-close/b256/step-002000/dataset_meta/modality.json +74 -0
  15. a7-cabinet-close/b256/step-002000/embodiment_id.json +57 -0
  16. a7-cabinet-close/b256/step-002000/experiment_cfg/conf.yaml +234 -0
  17. a7-cabinet-close/b256/step-002000/experiment_cfg/config.yaml +232 -0
  18. a7-cabinet-close/b256/step-002000/experiment_cfg/dataset_statistics.json +883 -0
  19. a7-cabinet-close/b256/step-002000/experiment_cfg/final_model_config.json +58 -0
  20. a7-cabinet-close/b256/step-002000/experiment_cfg/final_processor_config.json +0 -0
  21. a7-cabinet-close/b256/step-002000/model-00001-of-00002.safetensors +3 -0
  22. a7-cabinet-close/b256/step-002000/model-00002-of-00002.safetensors +3 -0
  23. a7-cabinet-close/b256/step-002000/model.safetensors.index.json +0 -0
  24. a7-cabinet-close/b256/step-002000/processor_config.json +1137 -0
  25. a7-cabinet-close/b256/step-002000/protocol.json +62 -0
  26. a7-cabinet-close/b256/step-002000/statistics.json +0 -0
  27. a7-cabinet-close/b256/step-002000/train.log +0 -0
  28. a7-cabinet-close/b256/step-002000/trainer_state.json +1234 -0
  29. a7-cabinet-close/b256/step-002000/training_args.bin +3 -0
  30. a7-cabinet-close/b256/step-002000/training_plan.json +1 -0
  31. a7-cabinet-close/b256/step-002000/wandb_config.json +1 -0
  32. a7-cabinet-close/b256/step-004000/code/hb_groot/cache_cosmos_reason2.py +60 -0
  33. a7-cabinet-close/b256/step-004000/code/hb_groot/check_reason2_offline.py +26 -0
  34. a7-cabinet-close/b256/step-004000/code/hb_groot/g1_feedback76_h16.py +27 -0
  35. a7-cabinet-close/b256/step-004000/code/hb_groot/groot_chain.py +139 -0
  36. a7-cabinet-close/b256/step-004000/code/hb_groot/h2_body31_hand44_h16.py +24 -0
  37. a7-cabinet-close/b256/step-004000/code/hb_groot/h2w3_body31_hand44_h16.py +24 -0
  38. a7-cabinet-close/b256/step-004000/code/hb_groot/hb_aug_v2.py +163 -0
  39. a7-cabinet-close/b256/step-004000/code/hb_groot/hb_launch.py +92 -0
  40. a7-cabinet-close/b256/step-004000/code/hb_groot/hb_offline_fix.py +24 -0
  41. a7-cabinet-close/b256/step-004000/code/hb_groot/prepare_groot_data.py +162 -0
  42. a7-cabinet-close/b256/step-004000/code/hb_groot/verify_groot_cpu.py +51 -0
  43. a7-cabinet-close/b256/step-004000/config.json +90 -0
  44. a7-cabinet-close/b256/step-004000/dataset_meta/hb-dataset-receipt.json +45 -0
  45. a7-cabinet-close/b256/step-004000/dataset_meta/modality.json +74 -0
  46. a7-cabinet-close/b256/step-004000/embodiment_id.json +57 -0
  47. a7-cabinet-close/b256/step-004000/experiment_cfg/conf.yaml +234 -0
  48. a7-cabinet-close/b256/step-004000/experiment_cfg/config.yaml +232 -0
  49. a7-cabinet-close/b256/step-004000/experiment_cfg/dataset_statistics.json +883 -0
  50. a7-cabinet-close/b256/step-004000/experiment_cfg/final_model_config.json +58 -0
a7-cabinet-close/b256/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)
a7-cabinet-close/b256/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)
a7-cabinet-close/b256/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)
a7-cabinet-close/b256/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")
a7-cabinet-close/b256/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)
a7-cabinet-close/b256/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)
a7-cabinet-close/b256/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)
a7-cabinet-close/b256/step-002000/code/hb_groot/hb_launch.py ADDED
@@ -0,0 +1,92 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ if os.environ.get("HB_PATCH_ONLY") != "1":
78
+ entry = Path(os.environ["GROOT_SRC"]) / "gr00t/experiment/launch_finetune.py"
79
+ sys.argv = [str(entry)] + sys.argv[1:]
80
+ try:
81
+ runpy.run_path(str(entry), run_name="__main__")
82
+ finally: # batch probe: report this rank's peak reserved memory (groot_chain.py compares it with MAX_FRAC)
83
+ _peak = os.environ.get("HB_PEAK_FILE")
84
+ if _peak and torch.cuda.is_available():
85
+ try:
86
+ _dev = torch.cuda.current_device()
87
+ with open(f"{_peak}-rank{os.environ.get('RANK', '0')}.json", "w") as _f:
88
+ json.dump({"rank": int(os.environ.get("RANK", "0")),
89
+ "peak_reserved_gb": torch.cuda.max_memory_reserved(_dev) / 1e9,
90
+ "total_gb": torch.cuda.get_device_properties(_dev).total_memory / 1e9}, _f)
91
+ except Exception: # noqa: BLE001 - never mask the training result
92
+ pass
a7-cabinet-close/b256/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
a7-cabinet-close/b256/step-002000/code/hb_groot/prepare_groot_data.py ADDED
@@ -0,0 +1,162 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "trolley": (P / "trolley-push-expert-v1-20261110/data/trolley_push", 100), # HF trolley-push-expert-v1 @ef9024a: fixed first 100 of 103
21
+ "b7": (P / "b7-black-clothes-hamper-v1-1bb002a/data", 100), # HF b7-black-clothes-hamper-expert-v1 @1bb002a, first 100 of 107
22
+ "a6": (P / "a6-door-pull-v1-cf3ee9b/data/door_pull", None), # HF door-pull-expert-v1 @cf3ee9b (A6)
23
+ "c6t": (P / "c6-test8-06c231d/data", None), # c6 8-episode adaptation test only
24
+ "c6": (P / "c6-bathroom-door-fetch-v1-3afba69/data", None), # HF c6-bathroom-door-fetch-expert-v1 @3afba69 (C6)
25
+ "a3": (P / "a3-a7-20261109/a3-fridge-open-expert-v1/780fcf1e4bd7db947a8c2d8f7b92e61f24f71b45", None),
26
+ "a7": (P / "a3-a7-20261109/a7-cabinet-close-expert-v1/1935a98e8f55d4d1d2d1615ddb97c2679713595c", None),
27
+ "a18": (P / "a18-h2-dataset-audit-8e956a10/dataset", None),
28
+ "a17": (P / "a17-unscrew-bottle-v1-407e6e2/data", None), # HF a17-unscrew-bottle-expert-v1 @407e6e2
29
+ "a14": (P / "a14-unscrew-bulb-v1-b4390fc/data", None), # HF a14-unscrew-bulb-expert-v1 @b4390fc (H2, 3 cameras)
30
+ }
31
+ CONFIG = {"a18": "h2w3_body31_hand44_h16.py", "a17": "h2w3_body31_hand44_h16.py", "a14": "h2w3_body31_hand44_h16.py"}
32
+ # model-input cameras per H2 task (user 2026-11-12 22:0x: EVERY H2 task uses head + chest + wrist; third_person never)
33
+ CAMS = {"a18": ("head", "chest", "wrist"), "a17": ("head", "chest", "wrist"), "a14": ("head", "chest", "wrist")}
34
+ H2_GROUPS = {"left_leg": (0, 6), "right_leg": (6, 12), "waist": (12, 15), "left_arm": (15, 22), "right_arm": (22, 29),
35
+ "head": (29, 31), "left_hand": (31, 53), "right_hand": (53, 75)}
36
+
37
+
38
+ def h2_modality(info, cams=("head", "chest")):
39
+ f = info["features"]
40
+ names = f["observation.state"]["names"]
41
+ assert len(names) == 75 and names[:31] == f["action.body_target"]["names"] and names[31:] == f["action.hands"]["names"]
42
+ assert f["observation.projected_gravity"]["shape"] == [3]
43
+ state = {g: {"original_key": "observation.state", "start": a, "end": b} for g, (a, b) in H2_GROUPS.items()}
44
+ state["projected_gravity"] = {"original_key": "observation.projected_gravity", "start": 0, "end": 3}
45
+ return {"state": state,
46
+ "action": {"body_target": {"original_key": "action.body_target", "start": 0, "end": 31},
47
+ "left_hand_target": {"original_key": "action.hands", "start": 0, "end": 22},
48
+ "right_hand_target": {"original_key": "action.hands", "start": 22, "end": 44}},
49
+ "video": {c: {"original_key": f"observation.images.{c}"} for c in cams},
50
+ "annotation": {"human.task_description": {"original_key": "task_index"}}}
51
+ BODY = ["left_hip_pitch_joint", "left_hip_roll_joint", "left_hip_yaw_joint", "left_knee_joint", "left_ankle_pitch_joint",
52
+ "left_ankle_roll_joint", "right_hip_pitch_joint", "right_hip_roll_joint", "right_hip_yaw_joint",
53
+ "right_knee_joint", "right_ankle_pitch_joint", "right_ankle_roll_joint", "waist_yaw_joint", "waist_roll_joint",
54
+ "waist_pitch_joint", "left_shoulder_pitch_joint", "left_shoulder_roll_joint", "left_shoulder_yaw_joint",
55
+ "left_elbow_joint", "left_wrist_roll_joint", "left_wrist_pitch_joint", "left_wrist_yaw_joint",
56
+ "right_shoulder_pitch_joint", "right_shoulder_roll_joint", "right_shoulder_yaw_joint", "right_elbow_joint",
57
+ "right_wrist_roll_joint", "right_wrist_pitch_joint", "right_wrist_yaw_joint"]
58
+ GROUPS = {"left_leg": (0, 6), "right_leg": (6, 12), "waist": (12, 15), "left_arm": (15, 22), "right_arm": (22, 29)}
59
+
60
+
61
+ def jl(path):
62
+ return [json.loads(x) for x in Path(path).read_text().splitlines() if x.strip()]
63
+
64
+
65
+ def wl(path, rows):
66
+ Path(path).write_text("".join(json.dumps(r) + "\n" for r in rows))
67
+
68
+
69
+ def build(task):
70
+ src, limit = TASKS[task]
71
+ info = json.loads((src / "meta/info.json").read_text())
72
+ names = info["features"]["observation.state"]["names"]
73
+ if task in CONFIG:
74
+ idx = list(range(31))
75
+ else:
76
+ idx = [names.index(n) for n in BODY]
77
+ for g, (a, b) in GROUPS.items():
78
+ assert idx[a:b] == list(range(idx[a], idx[a] + (b - a))), (task, g, idx[a:b])
79
+ for key, dim in (("action.motion_token", 64), ("observation.inspire_left_hand_state", 6),
80
+ ("observation.inspire_right_hand_state", 6), ("observation.projected_gravity", 3),
81
+ ("observation.images.ego_view", None), ("observation.images.chest", None)):
82
+ assert key in info["features"], (task, key)
83
+ if dim:
84
+ assert info["features"][key]["shape"] == [dim], (task, key)
85
+ episodes = jl(src / "meta/episodes.jsonl")
86
+ stats = jl(src / "meta/episodes_stats.jsonl")
87
+ n = len(episodes) if limit is None else limit
88
+ episodes, stats = episodes[:n], stats[:n]
89
+ assert [e["episode_index"] for e in episodes] == list(range(n)) == [s["episode_index"] for s in stats]
90
+ frames = sum(int(e["length"]) for e in episodes)
91
+ dst = OUT / task
92
+ if dst.exists():
93
+ raise SystemExit(f"{dst} exists")
94
+ (dst / "meta").mkdir(parents=True)
95
+ vkeys = [k for k, f in info["features"].items() if f.get("dtype") == "video"]
96
+ for e in episodes:
97
+ i = e["episode_index"]
98
+ chunk = i // int(info["chunks_size"])
99
+ rel = info["data_path"].format(episode_chunk=chunk, episode_index=i)
100
+ (dst / rel).parent.mkdir(parents=True, exist_ok=True)
101
+ os.symlink(src / rel, dst / rel)
102
+ for k in vkeys:
103
+ relv = info["video_path"].format(episode_chunk=chunk, video_key=k, episode_index=i)
104
+ (dst / relv).parent.mkdir(parents=True, exist_ok=True)
105
+ os.symlink(src / relv, dst / relv)
106
+ if task in CONFIG:
107
+ # stats.py computes stats for every low-dim feature in info.json and cannot handle e.g. observation.tactile
108
+ # [10,24,24]; keep only what the H2 modality uses (parquet columns stay untouched behind the symlinks).
109
+ keep = {"observation.state", "observation.projected_gravity", "action.body_target", "action.hands",
110
+ "timestamp", "frame_index", "episode_index", "index", "task_index"}
111
+ keep |= {f"observation.images.{c}" for c in CAMS[task]}
112
+ info["features"] = {k: v for k, v in info["features"].items() if k in keep}
113
+ vkeys = [k for k in vkeys if k in keep]
114
+ stats = [{**s, "stats": {k: v for k, v in s["stats"].items() if k in keep}} if "stats" in s else s
115
+ for s in stats]
116
+ info.update(total_episodes=n, total_frames=frames, total_videos=n * len(vkeys),
117
+ total_chunks=(n - 1) // int(info["chunks_size"]) + 1, splits={"train": f"0:{n}"})
118
+ (dst / "meta/info.json").write_text(json.dumps(info, indent=2))
119
+ wl(dst / "meta/episodes.jsonl", episodes)
120
+ wl(dst / "meta/episodes_stats.jsonl", stats)
121
+ (dst / "meta/tasks.jsonl").write_text((src / "meta/tasks.jsonl").read_text())
122
+ if task in CONFIG:
123
+ modality = h2_modality(info, CAMS[task])
124
+ else:
125
+ state = {g: {"original_key": "observation.state", "start": idx[a], "end": idx[a] + (b - a)}
126
+ for g, (a, b) in GROUPS.items()}
127
+ state.update(left_hand={"original_key": "observation.inspire_left_hand_state", "start": 0, "end": 6},
128
+ right_hand={"original_key": "observation.inspire_right_hand_state", "start": 0, "end": 6},
129
+ projected_gravity={"original_key": "observation.projected_gravity", "start": 0, "end": 3})
130
+ modality = {"state": state,
131
+ "action": {"motion_token": {"original_key": "action.motion_token", "start": 0, "end": 64},
132
+ "left_hand_joints": {"original_key": "observation.inspire_left_hand_state", "start": 0, "end": 6},
133
+ "right_hand_joints": {"original_key": "observation.inspire_right_hand_state", "start": 0, "end": 6}},
134
+ "video": {"ego_view": {"original_key": "observation.images.ego_view"},
135
+ "chest": {"original_key": "observation.images.chest"}},
136
+ "annotation": {"human.task_description": {"original_key": "task_index"}}}
137
+ (dst / "meta/modality.json").write_text(json.dumps(modality, indent=2))
138
+ py = sys.executable
139
+ src_g = G / "HumanWAM/third_party/Isaac-GR00T"
140
+ env = dict(os.environ, PYTHONPATH=f"{src_g}:{H}")
141
+ done = None
142
+ for tag in ("NEW_EMBODIMENT", "new_embodiment"):
143
+ r = subprocess.run([py, str(src_g / "gr00t/data/stats.py"), "--dataset-path", str(dst), "--embodiment-tag", tag,
144
+ "--modality-config-path", str(H / CONFIG.get(task, "g1_feedback76_h16.py"))], env=env, cwd=src_g,
145
+ capture_output=True, text=True)
146
+ if r.returncode == 0:
147
+ done = tag
148
+ break
149
+ last = r.stdout[-1500:] + r.stderr[-1500:]
150
+ if done is None:
151
+ raise SystemExit(f"stats failed for {task}:\n{last}")
152
+ assert (dst / "meta/stats.json").is_file()
153
+ receipt = dict(task=task, source=str(src), episodes=n, frames=frames, video_keys=vkeys,
154
+ state_dim_raw=len(names), body_index=idx, stats_tag=done,
155
+ modality_sha256=hashlib.sha256((dst / "meta/modality.json").read_bytes()).hexdigest(),
156
+ stats_sha256=hashlib.sha256((dst / "meta/stats.json").read_bytes()).hexdigest())
157
+ (dst / "hb-dataset-receipt.json").write_text(json.dumps(receipt, indent=2) + "\n")
158
+ print("DATASET_OK", json.dumps(receipt)[:400], flush=True)
159
+
160
+
161
+ for t in sys.argv[1:]:
162
+ build(t)
a7-cabinet-close/b256/step-002000/code/hb_groot/verify_groot_cpu.py ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 in ("a14", "a18", "a17"): # every H2 task: 3 cameras head + chest + wrist (user 2026-11-12)
20
+ import h2w3_body31_hand44_h16 as C # noqa: E402
21
+ cfg = C.h2w3_body31_hand44_h16
22
+ else:
23
+ import g1_feedback76_h16 as C # noqa: E402
24
+ cfg = C.feedback76_h16
25
+ loader = LeRobotEpisodeLoader(view, cfg)
26
+ df = loader[0]
27
+ cols = list(df.columns)
28
+ row = df.iloc[20]
29
+ out = {"episodes": len(loader), "columns": cols[:40], "rows_ep0": len(df)}
30
+ state_dims = {c: int(np.asarray(row[c]).size) for c in cols if c.startswith("state.")}
31
+ action_dims = {c: int(np.asarray(row[c]).size) for c in cols if c.startswith("action.")}
32
+ out.update(state_dims=state_dims, action_dims=action_dims, state_total=sum(state_dims.values()),
33
+ action_total=sum(action_dims.values()))
34
+ vids = [c for c in cols if c.startswith("video.")]
35
+ imgs = [np.asarray(row[c]) for c in vids]
36
+ out["video"] = {c: list(i.shape) for c, i in zip(vids, imgs)}
37
+ x = torch.stack([torch.from_numpy(i).permute(2, 0, 1).float().div(255) for i in imgs])[None]
38
+ rng = np.random.default_rng(0)
39
+ boosts = {}
40
+ for _ in range(60):
41
+ y, p = A.augment(x, rng)
42
+ boosts[p["boost"]] = boosts.get(p["boost"], 0) + 1
43
+ assert y.shape == x.shape and torch.isfinite(y).all() and float(y.min()) >= 0 and float(y.max()) <= 1
44
+ out["aug_boost_counts_60"] = boosts
45
+ out["mean_abs_change_last"] = round(float((y - x).abs().mean()), 4)
46
+ q = np.zeros((1, 29))
47
+ dq = A.perturb_joints(q, rng)
48
+ out["proprio_max_abs_deg"] = round(float(np.degrees(np.abs(dq).max())), 3)
49
+ out["patched"] = Gr00tN1d7Processor._get_vlm_inputs.__name__ == "_get_vlm_inputs" and \
50
+ Gr00tN1d7Processor.__call__.__module__ == "hb_launch"
51
+ print("CPU_VERIFY", json.dumps(out), flush=True)
a7-cabinet-close/b256/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": 0.2,
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
+ }
a7-cabinet-close/b256/step-002000/dataset_meta/hb-dataset-receipt.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "a7",
3
+ "source": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/a3-a7-20261109/a7-cabinet-close-expert-v1/1935a98e8f55d4d1d2d1615ddb97c2679713595c",
4
+ "episodes": 100,
5
+ "frames": 30664,
6
+ "video_keys": [
7
+ "observation.images.ego_view",
8
+ "observation.images.chest"
9
+ ],
10
+ "state_dim_raw": 41,
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
+ ],
42
+ "stats_tag": "NEW_EMBODIMENT",
43
+ "modality_sha256": "aea724f383cf0aa4d74349163ac73ec7b1f6f30cf5a66efce0a12c50dfabf823",
44
+ "stats_sha256": "169e38827bb311d7b150aee2f59db9604c7bb219af7612fcaa63cb71e859d210"
45
+ }
a7-cabinet-close/b256/step-002000/dataset_meta/modality.json ADDED
@@ -0,0 +1,74 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "left_hand": {
29
+ "original_key": "observation.inspire_left_hand_state",
30
+ "start": 0,
31
+ "end": 6
32
+ },
33
+ "right_hand": {
34
+ "original_key": "observation.inspire_right_hand_state",
35
+ "start": 0,
36
+ "end": 6
37
+ },
38
+ "projected_gravity": {
39
+ "original_key": "observation.projected_gravity",
40
+ "start": 0,
41
+ "end": 3
42
+ }
43
+ },
44
+ "action": {
45
+ "motion_token": {
46
+ "original_key": "action.motion_token",
47
+ "start": 0,
48
+ "end": 64
49
+ },
50
+ "left_hand_joints": {
51
+ "original_key": "observation.inspire_left_hand_state",
52
+ "start": 0,
53
+ "end": 6
54
+ },
55
+ "right_hand_joints": {
56
+ "original_key": "observation.inspire_right_hand_state",
57
+ "start": 0,
58
+ "end": 6
59
+ }
60
+ },
61
+ "video": {
62
+ "ego_view": {
63
+ "original_key": "observation.images.ego_view"
64
+ },
65
+ "chest": {
66
+ "original_key": "observation.images.chest"
67
+ }
68
+ },
69
+ "annotation": {
70
+ "human.task_description": {
71
+ "original_key": "task_index"
72
+ }
73
+ }
74
+ }
a7-cabinet-close/b256/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
+ }
a7-cabinet-close/b256/step-002000/experiment_cfg/conf.yaml ADDED
@@ -0,0 +1,234 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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: 0.2
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: 0.2
55
+ final_dropout: true
56
+ output_dim: 1024
57
+ interleave_self_attention: true
58
+ num_inference_timesteps: 4
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: 0.2
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/a7
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
+ - ego_view
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
+ - left_hand
99
+ - right_hand
100
+ - projected_gravity
101
+ sin_cos_embedding_keys: null
102
+ mean_std_embedding_keys: null
103
+ action_configs: null
104
+ action:
105
+ delta_indices:
106
+ - 0
107
+ - 1
108
+ - 2
109
+ - 3
110
+ - 4
111
+ - 5
112
+ - 6
113
+ - 7
114
+ - 8
115
+ - 9
116
+ - 10
117
+ - 11
118
+ - 12
119
+ - 13
120
+ - 14
121
+ - 15
122
+ modality_keys:
123
+ - motion_token
124
+ - left_hand_joints
125
+ - right_hand_joints
126
+ sin_cos_embedding_keys: null
127
+ mean_std_embedding_keys: null
128
+ action_configs:
129
+ - rep: ABSOLUTE
130
+ type: NON_EEF
131
+ format: DEFAULT
132
+ state_key: null
133
+ - rep: ABSOLUTE
134
+ type: NON_EEF
135
+ format: DEFAULT
136
+ state_key: null
137
+ - rep: ABSOLUTE
138
+ type: NON_EEF
139
+ format: DEFAULT
140
+ state_key: null
141
+ language:
142
+ delta_indices:
143
+ - 0
144
+ modality_keys:
145
+ - annotation.human.task_description
146
+ sin_cos_embedding_keys: null
147
+ mean_std_embedding_keys: null
148
+ action_configs: null
149
+ download_cache: false
150
+ shard_size: 1024
151
+ episode_sampling_rate: 0.1
152
+ num_shards_per_epoch: 100000
153
+ ds_weights_alpha: null
154
+ override_pretraining_statistics: true
155
+ mode: single_turn
156
+ random_chop: 0.0
157
+ mock_dataset_mode: false
158
+ shuffle: true
159
+ seed: 42
160
+ multiprocessing_context: fork
161
+ allow_padding: false
162
+ subsample_ratio: 1.0
163
+ loader: fast
164
+ fast_loader:
165
+ decode_threads: 2
166
+ ffmpeg_threads: 1
167
+ shuffle_buffer: 0
168
+ min_fill: 0
169
+ image_crop_size:
170
+ - 244
171
+ - 244
172
+ image_target_size:
173
+ - 224
174
+ - 224
175
+ training:
176
+ output_dir: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/groot-runs-ft4k/a7/production
177
+ experiment_name: null
178
+ max_steps: 4000
179
+ global_batch_size: 256
180
+ per_gpu_batch_size: null
181
+ gradient_accumulation_steps: 1
182
+ learning_rate: 0.0001
183
+ lr_scheduler_type: cosine
184
+ weight_decay: 1.0e-05
185
+ warmup_ratio: 0.05
186
+ warmup_steps: 0
187
+ max_grad_norm: 1.0
188
+ optim: adamw_torch
189
+ start_from_checkpoint: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/groot-n17-base
190
+ skip_weight_loading: false
191
+ tf32: true
192
+ fp16: false
193
+ bf16: true
194
+ eval_bf16: true
195
+ logging_steps: 10
196
+ save_steps: 2000
197
+ save_total_limit: 2
198
+ save_vl_model: false
199
+ save_only_model: false
200
+ resume_from_checkpoint: false
201
+ upload_checkpoints: false
202
+ upload_every: 1000
203
+ upload_last_n_checkpoints: 5
204
+ max_concurrent_uploads: 2
205
+ eval_strategy: 'no'
206
+ eval_steps: 500
207
+ eval_set_split_ratio: 0.1
208
+ eval_batch_size: 2
209
+ save_best_eval_metric_name: ''
210
+ save_best_eval_metric_greater_is_better: true
211
+ deepspeed_stage: 2
212
+ gradient_checkpointing: false
213
+ transformers_trust_remote_code: true
214
+ transformers_local_files_only: false
215
+ transformers_cache_dir: null
216
+ transformers_access_token: null
217
+ use_ddp: false
218
+ ddp_bucket_cap_mb: 100
219
+ num_gpus: 8
220
+ dataloader_num_workers: 8
221
+ remove_unused_columns: false
222
+ use_wandb: false
223
+ wandb_project: finetune-gr00t-n1d7
224
+ enable_profiling: false
225
+ max_retries: 3
226
+ assert_loss_less_than: null
227
+ add_rl_callback: false
228
+ enable_open_loop_eval: false
229
+ open_loop_eval_traj_ids:
230
+ - 0
231
+ open_loop_eval_steps_per_traj: 100
232
+ open_loop_eval_plot_indices: null
233
+ max_steps: 4000
234
+ save_steps: 2000
a7-cabinet-close/b256/step-002000/experiment_cfg/config.yaml ADDED
@@ -0,0 +1,232 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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: 0.2
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: 0.2
55
+ final_dropout: true
56
+ output_dim: 1024
57
+ interleave_self_attention: true
58
+ num_inference_timesteps: 4
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: 0.2
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/a7
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
+ - ego_view
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
+ - left_hand
99
+ - right_hand
100
+ - projected_gravity
101
+ sin_cos_embedding_keys: null
102
+ mean_std_embedding_keys: null
103
+ action_configs: null
104
+ action:
105
+ delta_indices:
106
+ - 0
107
+ - 1
108
+ - 2
109
+ - 3
110
+ - 4
111
+ - 5
112
+ - 6
113
+ - 7
114
+ - 8
115
+ - 9
116
+ - 10
117
+ - 11
118
+ - 12
119
+ - 13
120
+ - 14
121
+ - 15
122
+ modality_keys:
123
+ - motion_token
124
+ - left_hand_joints
125
+ - right_hand_joints
126
+ sin_cos_embedding_keys: null
127
+ mean_std_embedding_keys: null
128
+ action_configs:
129
+ - rep: absolute
130
+ type: non_eef
131
+ format: default
132
+ state_key: null
133
+ - rep: absolute
134
+ type: non_eef
135
+ format: default
136
+ state_key: null
137
+ - rep: absolute
138
+ type: non_eef
139
+ format: default
140
+ state_key: null
141
+ language:
142
+ delta_indices:
143
+ - 0
144
+ modality_keys:
145
+ - annotation.human.task_description
146
+ sin_cos_embedding_keys: null
147
+ mean_std_embedding_keys: null
148
+ action_configs: null
149
+ download_cache: false
150
+ shard_size: 1024
151
+ episode_sampling_rate: 0.1
152
+ num_shards_per_epoch: 100000
153
+ ds_weights_alpha: null
154
+ override_pretraining_statistics: true
155
+ mode: single_turn
156
+ random_chop: 0.0
157
+ mock_dataset_mode: false
158
+ shuffle: true
159
+ seed: 42
160
+ multiprocessing_context: fork
161
+ allow_padding: false
162
+ subsample_ratio: 1.0
163
+ loader: fast
164
+ fast_loader:
165
+ decode_threads: 2
166
+ ffmpeg_threads: 1
167
+ shuffle_buffer: 0
168
+ min_fill: 0
169
+ image_crop_size:
170
+ - 244
171
+ - 244
172
+ image_target_size:
173
+ - 224
174
+ - 224
175
+ training:
176
+ output_dir: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/groot-runs-ft4k/a7/production
177
+ experiment_name: null
178
+ max_steps: 4000
179
+ global_batch_size: 256
180
+ per_gpu_batch_size: null
181
+ gradient_accumulation_steps: 1
182
+ learning_rate: 0.0001
183
+ lr_scheduler_type: cosine
184
+ weight_decay: 1.0e-05
185
+ warmup_ratio: 0.05
186
+ warmup_steps: 0
187
+ max_grad_norm: 1.0
188
+ optim: adamw_torch
189
+ start_from_checkpoint: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/groot-n17-base
190
+ skip_weight_loading: false
191
+ tf32: true
192
+ fp16: false
193
+ bf16: true
194
+ eval_bf16: true
195
+ logging_steps: 10
196
+ save_steps: 2000
197
+ save_total_limit: 2
198
+ save_vl_model: false
199
+ save_only_model: false
200
+ resume_from_checkpoint: false
201
+ upload_checkpoints: false
202
+ upload_every: 1000
203
+ upload_last_n_checkpoints: 5
204
+ max_concurrent_uploads: 2
205
+ eval_strategy: 'no'
206
+ eval_steps: 500
207
+ eval_set_split_ratio: 0.1
208
+ eval_batch_size: 2
209
+ save_best_eval_metric_name: ''
210
+ save_best_eval_metric_greater_is_better: true
211
+ deepspeed_stage: 2
212
+ gradient_checkpointing: false
213
+ transformers_trust_remote_code: true
214
+ transformers_local_files_only: false
215
+ transformers_cache_dir: null
216
+ transformers_access_token: null
217
+ use_ddp: false
218
+ ddp_bucket_cap_mb: 100
219
+ num_gpus: 8
220
+ dataloader_num_workers: 8
221
+ remove_unused_columns: false
222
+ use_wandb: false
223
+ wandb_project: finetune-gr00t-n1d7
224
+ enable_profiling: false
225
+ max_retries: 3
226
+ assert_loss_less_than: null
227
+ add_rl_callback: false
228
+ enable_open_loop_eval: false
229
+ open_loop_eval_traj_ids:
230
+ - 0
231
+ open_loop_eval_steps_per_traj: 100
232
+ open_loop_eval_plot_indices: null
a7-cabinet-close/b256/step-002000/experiment_cfg/dataset_statistics.json ADDED
@@ -0,0 +1,883 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "new_embodiment": {
3
+ "state": {
4
+ "left_leg": {
5
+ "min": [
6
+ -0.9279626607894897,
7
+ -0.20401069521903992,
8
+ -0.5286000370979309,
9
+ -0.015552200376987457,
10
+ -0.4715784192085266,
11
+ -0.1929469108581543
12
+ ],
13
+ "max": [
14
+ 0.34397855401039124,
15
+ 0.31646499037742615,
16
+ 0.6314767599105835,
17
+ 1.2863916158676147,
18
+ 0.25699329376220703,
19
+ 0.18627917766571045
20
+ ],
21
+ "mean": [
22
+ -0.35561004281044006,
23
+ 0.037816599011421204,
24
+ 0.09485039860010147,
25
+ 0.5194756388664246,
26
+ -0.1980385035276413,
27
+ -0.006605133879929781
28
+ ],
29
+ "std": [
30
+ 0.28290456533432007,
31
+ 0.09233193844556808,
32
+ 0.19659870862960815,
33
+ 0.20531852543354034,
34
+ 0.08587498962879181,
35
+ 0.072576604783535
36
+ ],
37
+ "q01": [
38
+ -0.84334124147892,
39
+ -0.15387645348906517,
40
+ -0.3828186219930649,
41
+ 0.19774391919374465,
42
+ -0.4022857713699341,
43
+ -0.1577625685930252
44
+ ],
45
+ "q99": [
46
+ 0.19437677353620522,
47
+ 0.24248618379235268,
48
+ 0.45563414484262454,
49
+ 1.1226949226856227,
50
+ 0.00831889577209938,
51
+ 0.12483298838138578
52
+ ]
53
+ },
54
+ "right_leg": {
55
+ "min": [
56
+ -1.0070744752883911,
57
+ -0.3764217495918274,
58
+ -0.7747820019721985,
59
+ -0.07910795509815216,
60
+ -0.4874956011772156,
61
+ -0.21903759241104126
62
+ ],
63
+ "max": [
64
+ 0.6431071162223816,
65
+ 0.14184899628162384,
66
+ 0.18549136817455292,
67
+ 1.1398687362670898,
68
+ 0.33599478006362915,
69
+ 0.21929782629013062
70
+ ],
71
+ "mean": [
72
+ -0.3908240795135498,
73
+ -0.1153804287314415,
74
+ -0.2100031077861786,
75
+ 0.5305911898612976,
76
+ -0.18644081056118011,
77
+ -0.0369383879005909
78
+ ],
79
+ "std": [
80
+ 0.2948094308376312,
81
+ 0.09279008209705353,
82
+ 0.14734967052936554,
83
+ 0.20262227952480316,
84
+ 0.09397339075803757,
85
+ 0.07671832293272018
86
+ ],
87
+ "q01": [
88
+ -0.8900217366218567,
89
+ -0.3212410768866539,
90
+ -0.5801075077056885,
91
+ 0.16907327070832254,
92
+ -0.3691000935435295,
93
+ -0.17217106655240058
94
+ ],
95
+ "q99": [
96
+ 0.14047862306237197,
97
+ 0.06311478950083252,
98
+ 0.0607764326035973,
99
+ 0.9524855113029468,
100
+ 0.07608542144298523,
101
+ 0.15499507382512084
102
+ ]
103
+ },
104
+ "waist": {
105
+ "min": [
106
+ -0.07198242843151093,
107
+ -0.08106081932783127,
108
+ -0.09916144609451294
109
+ ],
110
+ "max": [
111
+ 0.19520391523838043,
112
+ 0.11672026664018631,
113
+ 0.11673737317323685
114
+ ],
115
+ "mean": [
116
+ 0.015539770945906639,
117
+ 0.013902479782700539,
118
+ 0.016449792310595512
119
+ ],
120
+ "std": [
121
+ 0.05012740567326546,
122
+ 0.035609565675258636,
123
+ 0.024248970672488213
124
+ ],
125
+ "q01": [
126
+ -0.050797894336283206,
127
+ -0.05091617964208126,
128
+ -0.048496404103934763
129
+ ],
130
+ "q99": [
131
+ 0.1525713104009628,
132
+ 0.08131846114993095,
133
+ 0.08361814886331549
134
+ ]
135
+ },
136
+ "left_arm": {
137
+ "min": [
138
+ 0.08239381015300751,
139
+ 0.015524790622293949,
140
+ -0.1121721938252449,
141
+ -0.4009973704814911,
142
+ -0.06022924929857254,
143
+ -0.1200670674443245,
144
+ -0.14843131601810455
145
+ ],
146
+ "max": [
147
+ 0.5195285677909851,
148
+ 0.22663770616054535,
149
+ 0.30364537239074707,
150
+ -0.12988236546516418,
151
+ 0.09782146662473679,
152
+ 0.08013442903757095,
153
+ 0.13720038533210754
154
+ ],
155
+ "mean": [
156
+ 0.20338064432144165,
157
+ 0.1432969570159912,
158
+ 0.03276205062866211,
159
+ -0.21513384580612183,
160
+ 0.004481739364564419,
161
+ 0.007482144050300121,
162
+ -0.02025328017771244
163
+ ],
164
+ "std": [
165
+ 0.03379302844405174,
166
+ 0.025766393169760704,
167
+ 0.035842087119817734,
168
+ 0.03536125645041466,
169
+ 0.012149316258728504,
170
+ 0.029369434341788292,
171
+ 0.037967316806316376
172
+ ],
173
+ "q01": [
174
+ 0.12020795546472073,
175
+ 0.07591435350477696,
176
+ -0.032209906913340094,
177
+ -0.3325797915458679,
178
+ -0.023349081706255673,
179
+ -0.07990536816418171,
180
+ -0.1029418321698904
181
+ ],
182
+ "q99": [
183
+ 0.32057825535535717,
184
+ 0.20605838119983666,
185
+ 0.15985203474760054,
186
+ -0.1695161056518555,
187
+ 0.04425690572708841,
188
+ 0.062308642081916304,
189
+ 0.0753356870263814
190
+ ]
191
+ },
192
+ "right_arm": {
193
+ "min": [
194
+ -1.307084083557129,
195
+ -0.42307642102241516,
196
+ -0.15481454133987427,
197
+ -0.590183675289154,
198
+ -0.09855499118566513,
199
+ -0.19881458580493927,
200
+ -0.2657955288887024
201
+ ],
202
+ "max": [
203
+ 0.7159279584884644,
204
+ 0.18037880957126617,
205
+ 0.3666689097881317,
206
+ 1.4362146854400635,
207
+ 0.14007999002933502,
208
+ 0.13200220465660095,
209
+ 0.18118087947368622
210
+ ],
211
+ "mean": [
212
+ -0.20058324933052063,
213
+ -0.07724551111459732,
214
+ 0.036659061908721924,
215
+ 0.21988210082054138,
216
+ 0.014881269074976444,
217
+ 0.009545939043164253,
218
+ 0.0062908814288675785
219
+ ],
220
+ "std": [
221
+ 0.47667089104652405,
222
+ 0.12800854444503784,
223
+ 0.08127856254577637,
224
+ 0.5201790928840637,
225
+ 0.031467556953430176,
226
+ 0.03061072714626789,
227
+ 0.08500789850950241
228
+ ],
229
+ "q01": [
230
+ -1.213955465555191,
231
+ -0.33947362542152404,
232
+ -0.08272466488182545,
233
+ -0.4007240226864815,
234
+ -0.06500550769269467,
235
+ -0.11041613429784775,
236
+ -0.21620410695672035
237
+ ],
238
+ "q99": [
239
+ 0.39352929234504636,
240
+ 0.16627973556518555,
241
+ 0.27784405529499046,
242
+ 1.3493560016155244,
243
+ 0.09599469408392904,
244
+ 0.07411817602813242,
245
+ 0.1364300513267517
246
+ ]
247
+ },
248
+ "left_hand": {
249
+ "min": [
250
+ 0.7841184139251709,
251
+ 0.9513298273086548,
252
+ 0.9574251770973206,
253
+ 0.9504732489585876,
254
+ 0.851092517375946,
255
+ 0.0
256
+ ],
257
+ "max": [
258
+ 1.0,
259
+ 1.0,
260
+ 1.0,
261
+ 1.0,
262
+ 1.0,
263
+ 0.30305954813957214
264
+ ],
265
+ "mean": [
266
+ 0.9994741678237915,
267
+ 0.9999215006828308,
268
+ 0.9999319911003113,
269
+ 0.9999361038208008,
270
+ 0.9997094869613647,
271
+ 0.0039056746754795313
272
+ ],
273
+ "std": [
274
+ 0.005472263786937151,
275
+ 0.0013332137605284705,
276
+ 0.0009734063060159317,
277
+ 0.000962596037425488,
278
+ 0.004112611990421562,
279
+ 0.028869736939668655
280
+ ],
281
+ "q01": [
282
+ 0.983048460483551,
283
+ 0.9994798117876053,
284
+ 0.9993437665700913,
285
+ 0.9995159441232682,
286
+ 0.9964931815862655,
287
+ 0.0
288
+ ],
289
+ "q99": [
290
+ 1.0,
291
+ 1.0,
292
+ 1.0,
293
+ 1.0,
294
+ 1.0,
295
+ 0.16704725056886674
296
+ ]
297
+ },
298
+ "right_hand": {
299
+ "min": [
300
+ 0.9825374484062195,
301
+ 0.9934524297714233,
302
+ 0.9735382199287415,
303
+ 0.9164777994155884,
304
+ 0.9381963014602661,
305
+ 0.0
306
+ ],
307
+ "max": [
308
+ 1.0,
309
+ 1.0,
310
+ 1.0,
311
+ 1.0,
312
+ 1.0,
313
+ 0.3030369281768799
314
+ ],
315
+ "mean": [
316
+ 0.9999874234199524,
317
+ 0.9999957084655762,
318
+ 0.9999594688415527,
319
+ 0.9998860359191895,
320
+ 0.9993945956230164,
321
+ 0.004138351418077946
322
+ ],
323
+ "std": [
324
+ 0.00026499829262562324,
325
+ 7.441091306000585e-05,
326
+ 0.0006803626310974737,
327
+ 0.001620433293278436,
328
+ 0.002877759747200986,
329
+ 0.0288485586643219
330
+ ],
331
+ "q01": [
332
+ 0.9998595207929611,
333
+ 0.9998556685447693,
334
+ 0.99947749376297,
335
+ 0.9987725448608399,
336
+ 0.9853247946500778,
337
+ 0.0
338
+ ],
339
+ "q99": [
340
+ 1.0,
341
+ 1.0,
342
+ 1.0,
343
+ 1.0,
344
+ 1.0,
345
+ 0.16710025072097778
346
+ ]
347
+ },
348
+ "projected_gravity": {
349
+ "min": [
350
+ -0.29140183329582214,
351
+ -0.18303944170475006,
352
+ -0.9999998807907104
353
+ ],
354
+ "max": [
355
+ 0.4053841233253479,
356
+ 0.10588564723730087,
357
+ -0.9137668013572693
358
+ ],
359
+ "mean": [
360
+ 0.08264563977718353,
361
+ -0.01804521679878235,
362
+ -0.986268162727356
363
+ ],
364
+ "std": [
365
+ 0.13568265736103058,
366
+ 0.0389891155064106,
367
+ 0.015453053638340727
368
+ ],
369
+ "q01": [
370
+ -0.13602852791547776,
371
+ -0.12972312435507774,
372
+ -0.999962568283081
373
+ ],
374
+ "q99": [
375
+ 0.325051891207695,
376
+ 0.0649579370766878,
377
+ -0.9452181661128998
378
+ ]
379
+ }
380
+ },
381
+ "action": {
382
+ "motion_token": {
383
+ "min": [
384
+ -0.0625,
385
+ -0.25,
386
+ -0.5625,
387
+ -0.375,
388
+ -0.3125,
389
+ -0.0625,
390
+ -0.0625,
391
+ -0.3125,
392
+ -0.125,
393
+ -0.3125,
394
+ -0.375,
395
+ -0.3125,
396
+ -0.375,
397
+ -0.125,
398
+ -0.0625,
399
+ -0.4375,
400
+ -0.1875,
401
+ -0.3125,
402
+ -0.4375,
403
+ -0.25,
404
+ -0.3125,
405
+ -0.4375,
406
+ -0.1875,
407
+ -0.1875,
408
+ 0.0,
409
+ -0.1875,
410
+ -0.3125,
411
+ -0.1875,
412
+ -0.1875,
413
+ -0.25,
414
+ -0.125,
415
+ -0.3125,
416
+ -0.3125,
417
+ -0.0625,
418
+ -0.1875,
419
+ -0.25,
420
+ -0.1875,
421
+ -0.375,
422
+ 0.0,
423
+ -0.1875,
424
+ -0.25,
425
+ -0.3125,
426
+ -0.125,
427
+ -0.25,
428
+ -0.125,
429
+ -0.3125,
430
+ -0.4375,
431
+ -0.5625,
432
+ -0.1875,
433
+ -0.5,
434
+ -0.25,
435
+ -0.1875,
436
+ -0.1875,
437
+ -0.3125,
438
+ -0.125,
439
+ -0.1875,
440
+ -0.4375,
441
+ -0.125,
442
+ -0.125,
443
+ -0.3125,
444
+ -0.25,
445
+ -0.1875,
446
+ -0.0625,
447
+ -0.25
448
+ ],
449
+ "max": [
450
+ 0.25,
451
+ 0.25,
452
+ 0.0625,
453
+ 0.3125,
454
+ 0.25,
455
+ 0.3125,
456
+ 0.3125,
457
+ 0.1875,
458
+ 0.1875,
459
+ -0.0625,
460
+ 0.125,
461
+ 0.125,
462
+ 0.125,
463
+ 0.25,
464
+ 0.5,
465
+ 0.25,
466
+ 0.1875,
467
+ 0.25,
468
+ 0.125,
469
+ 0.25,
470
+ 0.25,
471
+ 0.125,
472
+ 0.3125,
473
+ 0.4375,
474
+ 0.25,
475
+ 0.25,
476
+ 0.0,
477
+ 0.3125,
478
+ 0.375,
479
+ 0.1875,
480
+ 0.375,
481
+ 0.125,
482
+ 0.25,
483
+ 0.375,
484
+ 0.125,
485
+ 0.1875,
486
+ 0.4375,
487
+ 0.1875,
488
+ 0.4375,
489
+ 0.25,
490
+ 0.25,
491
+ 0.125,
492
+ 0.25,
493
+ 0.125,
494
+ 0.375,
495
+ 0.0625,
496
+ 0.1875,
497
+ 0.0625,
498
+ 0.3125,
499
+ 0.0625,
500
+ 0.375,
501
+ 0.25,
502
+ 0.5,
503
+ 0.0625,
504
+ 0.25,
505
+ 0.5,
506
+ 0.0625,
507
+ 0.3125,
508
+ 0.25,
509
+ 0.1875,
510
+ 0.25,
511
+ 0.1875,
512
+ 0.375,
513
+ 0.1875
514
+ ],
515
+ "mean": [
516
+ 0.1088450625538826,
517
+ 0.005287144333124161,
518
+ -0.19314587116241455,
519
+ 0.032281339168548584,
520
+ -0.026981966570019722,
521
+ 0.10573270171880722,
522
+ 0.15308871865272522,
523
+ -0.03123573213815689,
524
+ 0.004007142037153244,
525
+ -0.19649875164031982,
526
+ -0.06992931663990021,
527
+ -0.07310078293085098,
528
+ -0.10089192539453506,
529
+ 0.10953602194786072,
530
+ 0.1817685216665268,
531
+ -0.08574794232845306,
532
+ 0.05256571248173714,
533
+ -0.00392561312764883,
534
+ -0.16298019886016846,
535
+ 0.02755674347281456,
536
+ 0.009571484290063381,
537
+ -0.08721546083688736,
538
+ 0.03227726370096207,
539
+ 0.14658068120479584,
540
+ 0.09666057676076889,
541
+ 0.033968985080718994,
542
+ -0.14083494246006012,
543
+ 0.05588189885020256,
544
+ 0.10194160789251328,
545
+ -0.02117711305618286,
546
+ 0.1287054866552353,
547
+ -0.08698106557130814,
548
+ -0.0035730008967220783,
549
+ 0.15965178608894348,
550
+ -0.01520716492086649,
551
+ 0.007881799712777138,
552
+ 0.041449256241321564,
553
+ -0.08795329928398132,
554
+ 0.1488349586725235,
555
+ 0.010747537948191166,
556
+ 0.022936098277568817,
557
+ -0.09044400602579117,
558
+ 0.05614686757326126,
559
+ -0.1165250763297081,
560
+ 0.08106615394353867,
561
+ -0.10055765509605408,
562
+ -0.09149572998285294,
563
+ -0.21917599439620972,
564
+ 0.00674651050940156,
565
+ -0.15472948551177979,
566
+ 0.09517268091440201,
567
+ 0.0023908328730612993,
568
+ 0.08028755336999893,
569
+ -0.11982903629541397,
570
+ 0.09656070917844772,
571
+ 0.1366056352853775,
572
+ -0.13778983056545258,
573
+ 0.07862232625484467,
574
+ 0.08546259254217148,
575
+ -0.0442986898124218,
576
+ -0.052129533141851425,
577
+ 0.0008152882801368833,
578
+ 0.157947838306427,
579
+ -0.031631145626306534
580
+ ],
581
+ "std": [
582
+ 0.0668804869055748,
583
+ 0.12095443159341812,
584
+ 0.12842515110969543,
585
+ 0.14495190978050232,
586
+ 0.1207330971956253,
587
+ 0.06679413467645645,
588
+ 0.08728994429111481,
589
+ 0.12497754395008087,
590
+ 0.07046829164028168,
591
+ 0.04337522014975548,
592
+ 0.07340089976787567,
593
+ 0.08057640492916107,
594
+ 0.10612399876117706,
595
+ 0.05006479471921921,
596
+ 0.10516314208507538,
597
+ 0.14025066792964935,
598
+ 0.08458932489156723,
599
+ 0.09508704394102097,
600
+ 0.10922928899526596,
601
+ 0.11999057978391647,
602
+ 0.11393564939498901,
603
+ 0.08346482366323471,
604
+ 0.09646101295948029,
605
+ 0.11847557872533798,
606
+ 0.041381414979696274,
607
+ 0.11736250668764114,
608
+ 0.039891693741083145,
609
+ 0.10182952135801315,
610
+ 0.08461461961269379,
611
+ 0.09703998267650604,
612
+ 0.08291352540254593,
613
+ 0.07950283586978912,
614
+ 0.12207935005426407,
615
+ 0.07671988010406494,
616
+ 0.06723712384700775,
617
+ 0.0650196447968483,
618
+ 0.10957574844360352,
619
+ 0.12713782489299774,
620
+ 0.07227926701307297,
621
+ 0.11771422624588013,
622
+ 0.09277141839265823,
623
+ 0.06503032147884369,
624
+ 0.06486136466264725,
625
+ 0.0484318882226944,
626
+ 0.10573384165763855,
627
+ 0.05409308895468712,
628
+ 0.1513296663761139,
629
+ 0.13459795713424683,
630
+ 0.09658853709697723,
631
+ 0.11133205890655518,
632
+ 0.09778749197721481,
633
+ 0.12356267124414444,
634
+ 0.17755742371082306,
635
+ 0.04896029829978943,
636
+ 0.07157403975725174,
637
+ 0.1577388197183609,
638
+ 0.10338503122329712,
639
+ 0.11795450747013092,
640
+ 0.060382284224033356,
641
+ 0.11109492182731628,
642
+ 0.1049015000462532,
643
+ 0.06797655671834946,
644
+ 0.07634958624839783,
645
+ 0.07761601358652115
646
+ ],
647
+ "q01": [
648
+ 0.0,
649
+ -0.25,
650
+ -0.4375,
651
+ -0.25,
652
+ -0.25,
653
+ -0.0625,
654
+ -0.0625,
655
+ -0.25,
656
+ -0.125,
657
+ -0.25,
658
+ -0.25,
659
+ -0.25,
660
+ -0.3125,
661
+ -0.0625,
662
+ 0.0,
663
+ -0.375,
664
+ -0.125,
665
+ -0.1875,
666
+ -0.375,
667
+ -0.1875,
668
+ -0.25,
669
+ -0.3125,
670
+ -0.125,
671
+ -0.125,
672
+ 0.0,
673
+ -0.1875,
674
+ -0.25,
675
+ -0.125,
676
+ -0.125,
677
+ -0.1875,
678
+ -0.0625,
679
+ -0.25,
680
+ -0.25,
681
+ 0.0,
682
+ -0.125,
683
+ -0.1875,
684
+ -0.125,
685
+ -0.375,
686
+ 0.0,
687
+ -0.1875,
688
+ -0.1875,
689
+ -0.25,
690
+ -0.0625,
691
+ -0.1875,
692
+ -0.0625,
693
+ -0.25,
694
+ -0.375,
695
+ -0.5,
696
+ -0.1875,
697
+ -0.4375,
698
+ -0.1875,
699
+ -0.1875,
700
+ -0.1875,
701
+ -0.25,
702
+ -0.0625,
703
+ -0.125,
704
+ -0.375,
705
+ -0.125,
706
+ -0.0625,
707
+ -0.25,
708
+ -0.1875,
709
+ -0.125,
710
+ -0.0625,
711
+ -0.1875
712
+ ],
713
+ "q99": [
714
+ 0.25,
715
+ 0.25,
716
+ 0.0,
717
+ 0.3125,
718
+ 0.1875,
719
+ 0.25,
720
+ 0.3125,
721
+ 0.1875,
722
+ 0.125,
723
+ -0.125,
724
+ 0.0625,
725
+ 0.0625,
726
+ 0.0625,
727
+ 0.1875,
728
+ 0.4375,
729
+ 0.1875,
730
+ 0.1875,
731
+ 0.1875,
732
+ 0.0625,
733
+ 0.25,
734
+ 0.1875,
735
+ 0.0625,
736
+ 0.25,
737
+ 0.375,
738
+ 0.1875,
739
+ 0.25,
740
+ -0.0625,
741
+ 0.25,
742
+ 0.25,
743
+ 0.1875,
744
+ 0.3125,
745
+ 0.0625,
746
+ 0.25,
747
+ 0.3125,
748
+ 0.125,
749
+ 0.125,
750
+ 0.375,
751
+ 0.125,
752
+ 0.3125,
753
+ 0.25,
754
+ 0.1875,
755
+ 0.0625,
756
+ 0.1875,
757
+ 0.0,
758
+ 0.3125,
759
+ 0.0,
760
+ 0.125,
761
+ 0.0,
762
+ 0.1875,
763
+ 0.0625,
764
+ 0.3125,
765
+ 0.1875,
766
+ 0.4375,
767
+ 0.0,
768
+ 0.1875,
769
+ 0.4375,
770
+ 0.0,
771
+ 0.25,
772
+ 0.1875,
773
+ 0.1875,
774
+ 0.1875,
775
+ 0.125,
776
+ 0.3125,
777
+ 0.125
778
+ ]
779
+ },
780
+ "left_hand_joints": {
781
+ "min": [
782
+ 0.7841184139251709,
783
+ 0.9513298273086548,
784
+ 0.9574251770973206,
785
+ 0.9504732489585876,
786
+ 0.851092517375946,
787
+ 0.0
788
+ ],
789
+ "max": [
790
+ 1.0,
791
+ 1.0,
792
+ 1.0,
793
+ 1.0,
794
+ 1.0,
795
+ 0.30305954813957214
796
+ ],
797
+ "mean": [
798
+ 0.9994741678237915,
799
+ 0.9999215006828308,
800
+ 0.9999319911003113,
801
+ 0.9999361038208008,
802
+ 0.9997094869613647,
803
+ 0.0039056746754795313
804
+ ],
805
+ "std": [
806
+ 0.005472263786937151,
807
+ 0.0013332137605284705,
808
+ 0.0009734063060159317,
809
+ 0.000962596037425488,
810
+ 0.004112611990421562,
811
+ 0.028869736939668655
812
+ ],
813
+ "q01": [
814
+ 0.983048460483551,
815
+ 0.9994798117876053,
816
+ 0.9993437665700913,
817
+ 0.9995159441232682,
818
+ 0.9964931815862655,
819
+ 0.0
820
+ ],
821
+ "q99": [
822
+ 1.0,
823
+ 1.0,
824
+ 1.0,
825
+ 1.0,
826
+ 1.0,
827
+ 0.16704725056886674
828
+ ]
829
+ },
830
+ "right_hand_joints": {
831
+ "min": [
832
+ 0.9825374484062195,
833
+ 0.9934524297714233,
834
+ 0.9735382199287415,
835
+ 0.9164777994155884,
836
+ 0.9381963014602661,
837
+ 0.0
838
+ ],
839
+ "max": [
840
+ 1.0,
841
+ 1.0,
842
+ 1.0,
843
+ 1.0,
844
+ 1.0,
845
+ 0.3030369281768799
846
+ ],
847
+ "mean": [
848
+ 0.9999874234199524,
849
+ 0.9999957084655762,
850
+ 0.9999594688415527,
851
+ 0.9998860359191895,
852
+ 0.9993945956230164,
853
+ 0.004138351418077946
854
+ ],
855
+ "std": [
856
+ 0.00026499829262562324,
857
+ 7.441091306000585e-05,
858
+ 0.0006803626310974737,
859
+ 0.001620433293278436,
860
+ 0.002877759747200986,
861
+ 0.0288485586643219
862
+ ],
863
+ "q01": [
864
+ 0.9998595207929611,
865
+ 0.9998556685447693,
866
+ 0.99947749376297,
867
+ 0.9987725448608399,
868
+ 0.9853247946500778,
869
+ 0.0
870
+ ],
871
+ "q99": [
872
+ 1.0,
873
+ 1.0,
874
+ 1.0,
875
+ 1.0,
876
+ 1.0,
877
+ 0.16710025072097778
878
+ ]
879
+ }
880
+ },
881
+ "relative_action": {}
882
+ }
883
+ }
a7-cabinet-close/b256/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": 0.2,
55
+ "exclude_state": false,
56
+ "use_mean_std": false,
57
+ "max_num_embodiments": 32
58
+ }
a7-cabinet-close/b256/step-002000/experiment_cfg/final_processor_config.json ADDED
The diff for this file is too large to render. See raw diff
 
a7-cabinet-close/b256/step-002000/model-00001-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1f846fe9d000a8a3dd95a70bced02d79e65133b428d37282b6a2cb2f74bbe04e
3
+ size 4990519232
a7-cabinet-close/b256/step-002000/model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:baf37ac5beb3cad3752d58fbc6edd28392c4e69ff637f81735c9b64c7cc834b7
3
+ size 1919980184
a7-cabinet-close/b256/step-002000/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
a7-cabinet-close/b256/step-002000/processor_config.json ADDED
@@ -0,0 +1,1137 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "ego_view",
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
+ "left_hand",
1036
+ "right_hand",
1037
+ "projected_gravity"
1038
+ ],
1039
+ "sin_cos_embedding_keys": null,
1040
+ "mean_std_embedding_keys": null,
1041
+ "action_configs": null
1042
+ },
1043
+ "action": {
1044
+ "delta_indices": [
1045
+ 0,
1046
+ 1,
1047
+ 2,
1048
+ 3,
1049
+ 4,
1050
+ 5,
1051
+ 6,
1052
+ 7,
1053
+ 8,
1054
+ 9,
1055
+ 10,
1056
+ 11,
1057
+ 12,
1058
+ 13,
1059
+ 14,
1060
+ 15
1061
+ ],
1062
+ "modality_keys": [
1063
+ "motion_token",
1064
+ "left_hand_joints",
1065
+ "right_hand_joints"
1066
+ ],
1067
+ "sin_cos_embedding_keys": null,
1068
+ "mean_std_embedding_keys": null,
1069
+ "action_configs": [
1070
+ {
1071
+ "rep": "ABSOLUTE",
1072
+ "type": "NON_EEF",
1073
+ "format": "DEFAULT",
1074
+ "state_key": null
1075
+ },
1076
+ {
1077
+ "rep": "ABSOLUTE",
1078
+ "type": "NON_EEF",
1079
+ "format": "DEFAULT",
1080
+ "state_key": null
1081
+ },
1082
+ {
1083
+ "rep": "ABSOLUTE",
1084
+ "type": "NON_EEF",
1085
+ "format": "DEFAULT",
1086
+ "state_key": null
1087
+ }
1088
+ ]
1089
+ },
1090
+ "language": {
1091
+ "delta_indices": [
1092
+ 0
1093
+ ],
1094
+ "modality_keys": [
1095
+ "annotation.human.task_description"
1096
+ ],
1097
+ "sin_cos_embedding_keys": null,
1098
+ "mean_std_embedding_keys": null,
1099
+ "action_configs": null
1100
+ }
1101
+ }
1102
+ },
1103
+ "image_crop_size": [
1104
+ 230,
1105
+ 230
1106
+ ],
1107
+ "image_target_size": [
1108
+ 256,
1109
+ 256
1110
+ ],
1111
+ "use_albumentations": true,
1112
+ "random_rotation_angle": 0,
1113
+ "color_jitter_params": {
1114
+ "brightness": 0.3,
1115
+ "contrast": 0.4,
1116
+ "saturation": 0.5,
1117
+ "hue": 0.08
1118
+ },
1119
+ "shortest_image_edge": 256,
1120
+ "crop_fraction": 0.95,
1121
+ "letter_box_transform": false,
1122
+ "model_name": "nvidia/Cosmos-Reason2-2B",
1123
+ "model_type": "qwen",
1124
+ "formalize_language": true,
1125
+ "max_state_dim": 132,
1126
+ "max_action_dim": 132,
1127
+ "max_action_horizon": 40,
1128
+ "strict_action_padding_mask": false,
1129
+ "use_percentiles": true,
1130
+ "use_mean_std": false,
1131
+ "clip_outliers": true,
1132
+ "apply_sincos_state_encoding": false,
1133
+ "use_relative_action": true,
1134
+ "exclude_state": false,
1135
+ "state_dropout_prob": 0.2
1136
+ }
1137
+ }
a7-cabinet-close/b256/step-002000/protocol.json ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "GR00T N1.7 (Gr00tN1d7, action head fine-tuned, VLM frozen)",
3
+ "task": "A7 close cabinet door (G1)",
4
+ "task_dir": "a7-cabinet-close/b256",
5
+ "robot": "Unitree G1",
6
+ "training_data": "a7-cabinet-close-expert-v1",
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": 256,
15
+ "per_gpu_batch": 32,
16
+ "num_gpus": 8,
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
+ "epochs_equivalent": 33.39,
25
+ "recipe": "full fine-tune (tune_llm, tune_visual, projector, diffusion, vlln), fixed steps",
26
+ "gradient_accumulation_steps": 1,
27
+ "trainable": "all modules (tune_llm, tune_visual, projector, diffusion action head, VLLN)",
28
+ "training_records_included": true,
29
+ "resume_state_included": false,
30
+ "files": {
31
+ "code/hb_groot/cache_cosmos_reason2.py": 2542,
32
+ "code/hb_groot/check_reason2_offline.py": 1283,
33
+ "code/hb_groot/g1_feedback76_h16.py": 1474,
34
+ "code/hb_groot/groot_chain.py": 6325,
35
+ "code/hb_groot/h2_body31_hand44_h16.py": 1395,
36
+ "code/hb_groot/h2w3_body31_hand44_h16.py": 1563,
37
+ "code/hb_groot/hb_aug_v2.py": 7475,
38
+ "code/hb_groot/hb_launch.py": 3913,
39
+ "code/hb_groot/hb_offline_fix.py": 1138,
40
+ "code/hb_groot/prepare_groot_data.py": 10239,
41
+ "code/hb_groot/verify_groot_cpu.py": 2310,
42
+ "config.json": 2219,
43
+ "dataset_meta/hb-dataset-receipt.json": 817,
44
+ "dataset_meta/modality.json": 1537,
45
+ "embodiment_id.json": 2226,
46
+ "experiment_cfg/conf.yaml": 5792,
47
+ "experiment_cfg/config.yaml": 5759,
48
+ "experiment_cfg/dataset_statistics.json": 20397,
49
+ "experiment_cfg/final_model_config.json": 1584,
50
+ "experiment_cfg/final_processor_config.json": 2548329,
51
+ "model-00001-of-00002.safetensors": 4990519232,
52
+ "model-00002-of-00002.safetensors": 1919980184,
53
+ "model.safetensors.index.json": 104985,
54
+ "processor_config.json": 26945,
55
+ "statistics.json": 4313490,
56
+ "train.log": 385943,
57
+ "trainer_state.json": 28270,
58
+ "training_args.bin": 8017,
59
+ "training_plan.json": 285,
60
+ "wandb_config.json": 58
61
+ }
62
+ }
a7-cabinet-close/b256/step-002000/statistics.json ADDED
The diff for this file is too large to render. See raw diff
 
a7-cabinet-close/b256/step-002000/train.log ADDED
The diff for this file is too large to render. See raw diff
 
a7-cabinet-close/b256/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": 0.3687867522239685,
14
+ "learning_rate": 4.5e-06,
15
+ "loss": 1.3585,
16
+ "step": 10
17
+ },
18
+ {
19
+ "grad_norm": 0.6812222003936768,
20
+ "learning_rate": 9.5e-06,
21
+ "loss": 1.2938,
22
+ "step": 20
23
+ },
24
+ {
25
+ "grad_norm": 0.4920693337917328,
26
+ "learning_rate": 1.45e-05,
27
+ "loss": 1.1866,
28
+ "step": 30
29
+ },
30
+ {
31
+ "grad_norm": 0.3045933246612549,
32
+ "learning_rate": 1.9500000000000003e-05,
33
+ "loss": 1.1416,
34
+ "step": 40
35
+ },
36
+ {
37
+ "grad_norm": 0.3579404950141907,
38
+ "learning_rate": 2.45e-05,
39
+ "loss": 1.1155,
40
+ "step": 50
41
+ },
42
+ {
43
+ "grad_norm": 0.27957281470298767,
44
+ "learning_rate": 2.95e-05,
45
+ "loss": 1.1142,
46
+ "step": 60
47
+ },
48
+ {
49
+ "grad_norm": 0.35014021396636963,
50
+ "learning_rate": 3.45e-05,
51
+ "loss": 1.1045,
52
+ "step": 70
53
+ },
54
+ {
55
+ "grad_norm": 0.34105372428894043,
56
+ "learning_rate": 3.9500000000000005e-05,
57
+ "loss": 1.1001,
58
+ "step": 80
59
+ },
60
+ {
61
+ "grad_norm": 0.530993640422821,
62
+ "learning_rate": 4.4500000000000004e-05,
63
+ "loss": 1.0875,
64
+ "step": 90
65
+ },
66
+ {
67
+ "grad_norm": 0.5785247683525085,
68
+ "learning_rate": 4.9500000000000004e-05,
69
+ "loss": 1.045,
70
+ "step": 100
71
+ },
72
+ {
73
+ "grad_norm": 0.776517927646637,
74
+ "learning_rate": 5.45e-05,
75
+ "loss": 0.9913,
76
+ "step": 110
77
+ },
78
+ {
79
+ "grad_norm": 0.8615429401397705,
80
+ "learning_rate": 5.95e-05,
81
+ "loss": 0.9325,
82
+ "step": 120
83
+ },
84
+ {
85
+ "grad_norm": 0.6859230399131775,
86
+ "learning_rate": 6.450000000000001e-05,
87
+ "loss": 0.8829,
88
+ "step": 130
89
+ },
90
+ {
91
+ "grad_norm": 0.8569195866584778,
92
+ "learning_rate": 6.95e-05,
93
+ "loss": 0.8446,
94
+ "step": 140
95
+ },
96
+ {
97
+ "grad_norm": 0.9444453120231628,
98
+ "learning_rate": 7.450000000000001e-05,
99
+ "loss": 0.8069,
100
+ "step": 150
101
+ },
102
+ {
103
+ "grad_norm": 0.7435154914855957,
104
+ "learning_rate": 7.950000000000001e-05,
105
+ "loss": 0.7831,
106
+ "step": 160
107
+ },
108
+ {
109
+ "grad_norm": 0.8777171969413757,
110
+ "learning_rate": 8.450000000000001e-05,
111
+ "loss": 0.7521,
112
+ "step": 170
113
+ },
114
+ {
115
+ "grad_norm": 1.0203330516815186,
116
+ "learning_rate": 8.950000000000001e-05,
117
+ "loss": 0.746,
118
+ "step": 180
119
+ },
120
+ {
121
+ "grad_norm": 0.8279154300689697,
122
+ "learning_rate": 9.449999999999999e-05,
123
+ "loss": 0.7197,
124
+ "step": 190
125
+ },
126
+ {
127
+ "grad_norm": 0.8975831866264343,
128
+ "learning_rate": 9.95e-05,
129
+ "loss": 0.713,
130
+ "step": 200
131
+ },
132
+ {
133
+ "grad_norm": 0.7117353677749634,
134
+ "learning_rate": 9.999861593790126e-05,
135
+ "loss": 0.6976,
136
+ "step": 210
137
+ },
138
+ {
139
+ "grad_norm": 0.9157130718231201,
140
+ "learning_rate": 9.999383162408304e-05,
141
+ "loss": 0.6747,
142
+ "step": 220
143
+ },
144
+ {
145
+ "grad_norm": 0.9195032715797424,
146
+ "learning_rate": 9.998563029828259e-05,
147
+ "loss": 0.6566,
148
+ "step": 230
149
+ },
150
+ {
151
+ "grad_norm": 1.155577301979065,
152
+ "learning_rate": 9.997401252104962e-05,
153
+ "loss": 0.6317,
154
+ "step": 240
155
+ },
156
+ {
157
+ "grad_norm": 0.9686551094055176,
158
+ "learning_rate": 9.995897908644378e-05,
159
+ "loss": 0.608,
160
+ "step": 250
161
+ },
162
+ {
163
+ "grad_norm": 1.1770495176315308,
164
+ "learning_rate": 9.994053102198034e-05,
165
+ "loss": 0.5837,
166
+ "step": 260
167
+ },
168
+ {
169
+ "grad_norm": 0.9421539902687073,
170
+ "learning_rate": 9.991866958856003e-05,
171
+ "loss": 0.5745,
172
+ "step": 270
173
+ },
174
+ {
175
+ "grad_norm": 1.036603331565857,
176
+ "learning_rate": 9.989339628038276e-05,
177
+ "loss": 0.5418,
178
+ "step": 280
179
+ },
180
+ {
181
+ "grad_norm": 1.00693678855896,
182
+ "learning_rate": 9.98647128248456e-05,
183
+ "loss": 0.5354,
184
+ "step": 290
185
+ },
186
+ {
187
+ "grad_norm": 0.9830642342567444,
188
+ "learning_rate": 9.98326211824246e-05,
189
+ "loss": 0.5092,
190
+ "step": 300
191
+ },
192
+ {
193
+ "grad_norm": 1.0696625709533691,
194
+ "learning_rate": 9.979712354654091e-05,
195
+ "loss": 0.4975,
196
+ "step": 310
197
+ },
198
+ {
199
+ "grad_norm": 0.9277350902557373,
200
+ "learning_rate": 9.975822234341079e-05,
201
+ "loss": 0.4685,
202
+ "step": 320
203
+ },
204
+ {
205
+ "grad_norm": 1.228633165359497,
206
+ "learning_rate": 9.97159202318798e-05,
207
+ "loss": 0.4607,
208
+ "step": 330
209
+ },
210
+ {
211
+ "grad_norm": 0.9199367165565491,
212
+ "learning_rate": 9.967022010324105e-05,
213
+ "loss": 0.4421,
214
+ "step": 340
215
+ },
216
+ {
217
+ "grad_norm": 1.2343778610229492,
218
+ "learning_rate": 9.962112508103765e-05,
219
+ "loss": 0.4217,
220
+ "step": 350
221
+ },
222
+ {
223
+ "grad_norm": 0.8779021501541138,
224
+ "learning_rate": 9.956863852084914e-05,
225
+ "loss": 0.4132,
226
+ "step": 360
227
+ },
228
+ {
229
+ "grad_norm": 1.0454401969909668,
230
+ "learning_rate": 9.951276401006221e-05,
231
+ "loss": 0.4105,
232
+ "step": 370
233
+ },
234
+ {
235
+ "grad_norm": 1.1825932264328003,
236
+ "learning_rate": 9.945350536762543e-05,
237
+ "loss": 0.3878,
238
+ "step": 380
239
+ },
240
+ {
241
+ "grad_norm": 1.1252363920211792,
242
+ "learning_rate": 9.939086664378829e-05,
243
+ "loss": 0.3766,
244
+ "step": 390
245
+ },
246
+ {
247
+ "grad_norm": 1.0986732244491577,
248
+ "learning_rate": 9.932485211982437e-05,
249
+ "loss": 0.359,
250
+ "step": 400
251
+ },
252
+ {
253
+ "grad_norm": 1.228696584701538,
254
+ "learning_rate": 9.92554663077387e-05,
255
+ "loss": 0.3534,
256
+ "step": 410
257
+ },
258
+ {
259
+ "grad_norm": 1.1085554361343384,
260
+ "learning_rate": 9.918271394995935e-05,
261
+ "loss": 0.3354,
262
+ "step": 420
263
+ },
264
+ {
265
+ "grad_norm": 1.1127831935882568,
266
+ "learning_rate": 9.910660001901335e-05,
267
+ "loss": 0.3272,
268
+ "step": 430
269
+ },
270
+ {
271
+ "grad_norm": 1.0754750967025757,
272
+ "learning_rate": 9.902712971718675e-05,
273
+ "loss": 0.3085,
274
+ "step": 440
275
+ },
276
+ {
277
+ "grad_norm": 1.0768864154815674,
278
+ "learning_rate": 9.894430847616915e-05,
279
+ "loss": 0.3,
280
+ "step": 450
281
+ },
282
+ {
283
+ "grad_norm": 1.3451212644577026,
284
+ "learning_rate": 9.885814195668232e-05,
285
+ "loss": 0.2932,
286
+ "step": 460
287
+ },
288
+ {
289
+ "grad_norm": 1.4541419744491577,
290
+ "learning_rate": 9.876863604809344e-05,
291
+ "loss": 0.2847,
292
+ "step": 470
293
+ },
294
+ {
295
+ "grad_norm": 1.3149020671844482,
296
+ "learning_rate": 9.867579686801245e-05,
297
+ "loss": 0.2643,
298
+ "step": 480
299
+ },
300
+ {
301
+ "grad_norm": 1.0394401550292969,
302
+ "learning_rate": 9.8579630761874e-05,
303
+ "loss": 0.2491,
304
+ "step": 490
305
+ },
306
+ {
307
+ "grad_norm": 1.3146859407424927,
308
+ "learning_rate": 9.848014430250367e-05,
309
+ "loss": 0.2369,
310
+ "step": 500
311
+ },
312
+ {
313
+ "grad_norm": 1.4528263807296753,
314
+ "learning_rate": 9.837734428966885e-05,
315
+ "loss": 0.233,
316
+ "step": 510
317
+ },
318
+ {
319
+ "grad_norm": 1.1015030145645142,
320
+ "learning_rate": 9.827123774961383e-05,
321
+ "loss": 0.2183,
322
+ "step": 520
323
+ },
324
+ {
325
+ "grad_norm": 1.4509450197219849,
326
+ "learning_rate": 9.816183193457968e-05,
327
+ "loss": 0.2202,
328
+ "step": 530
329
+ },
330
+ {
331
+ "grad_norm": 0.9684346914291382,
332
+ "learning_rate": 9.804913432230856e-05,
333
+ "loss": 0.2013,
334
+ "step": 540
335
+ },
336
+ {
337
+ "grad_norm": 1.2104988098144531,
338
+ "learning_rate": 9.793315261553252e-05,
339
+ "loss": 0.1977,
340
+ "step": 550
341
+ },
342
+ {
343
+ "grad_norm": 1.2478097677230835,
344
+ "learning_rate": 9.781389474144717e-05,
345
+ "loss": 0.1804,
346
+ "step": 560
347
+ },
348
+ {
349
+ "grad_norm": 1.0177229642868042,
350
+ "learning_rate": 9.76913688511698e-05,
351
+ "loss": 0.1746,
352
+ "step": 570
353
+ },
354
+ {
355
+ "grad_norm": 1.6421438455581665,
356
+ "learning_rate": 9.756558331918227e-05,
357
+ "loss": 0.1725,
358
+ "step": 580
359
+ },
360
+ {
361
+ "grad_norm": 1.4186254739761353,
362
+ "learning_rate": 9.743654674275855e-05,
363
+ "loss": 0.1773,
364
+ "step": 590
365
+ },
366
+ {
367
+ "grad_norm": 1.2849631309509277,
368
+ "learning_rate": 9.730426794137727e-05,
369
+ "loss": 0.1714,
370
+ "step": 600
371
+ },
372
+ {
373
+ "grad_norm": 1.0712231397628784,
374
+ "learning_rate": 9.716875595611879e-05,
375
+ "loss": 0.1857,
376
+ "step": 610
377
+ },
378
+ {
379
+ "grad_norm": 1.0974456071853638,
380
+ "learning_rate": 9.703002004904729e-05,
381
+ "loss": 0.1547,
382
+ "step": 620
383
+ },
384
+ {
385
+ "grad_norm": 1.2957112789154053,
386
+ "learning_rate": 9.688806970257773e-05,
387
+ "loss": 0.1409,
388
+ "step": 630
389
+ },
390
+ {
391
+ "grad_norm": 1.255246877670288,
392
+ "learning_rate": 9.674291461882774e-05,
393
+ "loss": 0.1508,
394
+ "step": 640
395
+ },
396
+ {
397
+ "grad_norm": 1.2222325801849365,
398
+ "learning_rate": 9.659456471895445e-05,
399
+ "loss": 0.159,
400
+ "step": 650
401
+ },
402
+ {
403
+ "grad_norm": 1.0446699857711792,
404
+ "learning_rate": 9.644303014247648e-05,
405
+ "loss": 0.1528,
406
+ "step": 660
407
+ },
408
+ {
409
+ "grad_norm": 1.1961761713027954,
410
+ "learning_rate": 9.628832124658085e-05,
411
+ "loss": 0.1757,
412
+ "step": 670
413
+ },
414
+ {
415
+ "grad_norm": 1.083086609840393,
416
+ "learning_rate": 9.613044860541507e-05,
417
+ "loss": 0.1577,
418
+ "step": 680
419
+ },
420
+ {
421
+ "grad_norm": 1.0953060388565063,
422
+ "learning_rate": 9.596942300936445e-05,
423
+ "loss": 0.1492,
424
+ "step": 690
425
+ },
426
+ {
427
+ "grad_norm": 1.0373493432998657,
428
+ "learning_rate": 9.580525546431459e-05,
429
+ "loss": 0.1528,
430
+ "step": 700
431
+ },
432
+ {
433
+ "grad_norm": 1.0717378854751587,
434
+ "learning_rate": 9.563795719089911e-05,
435
+ "loss": 0.1665,
436
+ "step": 710
437
+ },
438
+ {
439
+ "grad_norm": 1.2267723083496094,
440
+ "learning_rate": 9.546753962373281e-05,
441
+ "loss": 0.1497,
442
+ "step": 720
443
+ },
444
+ {
445
+ "grad_norm": 0.9240042567253113,
446
+ "learning_rate": 9.529401441062997e-05,
447
+ "loss": 0.1518,
448
+ "step": 730
449
+ },
450
+ {
451
+ "grad_norm": 1.1921186447143555,
452
+ "learning_rate": 9.511739341180842e-05,
453
+ "loss": 0.1244,
454
+ "step": 740
455
+ },
456
+ {
457
+ "grad_norm": 1.0847467184066772,
458
+ "learning_rate": 9.493768869907886e-05,
459
+ "loss": 0.1489,
460
+ "step": 750
461
+ },
462
+ {
463
+ "grad_norm": 1.1856486797332764,
464
+ "learning_rate": 9.475491255501968e-05,
465
+ "loss": 0.1296,
466
+ "step": 760
467
+ },
468
+ {
469
+ "grad_norm": 0.9667167067527771,
470
+ "learning_rate": 9.456907747213748e-05,
471
+ "loss": 0.1392,
472
+ "step": 770
473
+ },
474
+ {
475
+ "grad_norm": 1.244463324546814,
476
+ "learning_rate": 9.438019615201336e-05,
477
+ "loss": 0.1382,
478
+ "step": 780
479
+ },
480
+ {
481
+ "grad_norm": 1.066081166267395,
482
+ "learning_rate": 9.418828150443469e-05,
483
+ "loss": 0.1321,
484
+ "step": 790
485
+ },
486
+ {
487
+ "grad_norm": 0.947472095489502,
488
+ "learning_rate": 9.399334664651262e-05,
489
+ "loss": 0.1365,
490
+ "step": 800
491
+ },
492
+ {
493
+ "grad_norm": 1.2648314237594604,
494
+ "learning_rate": 9.379540490178581e-05,
495
+ "loss": 0.1462,
496
+ "step": 810
497
+ },
498
+ {
499
+ "grad_norm": 1.1162383556365967,
500
+ "learning_rate": 9.359446979930955e-05,
501
+ "loss": 0.139,
502
+ "step": 820
503
+ },
504
+ {
505
+ "grad_norm": 1.0630543231964111,
506
+ "learning_rate": 9.33905550727312e-05,
507
+ "loss": 0.14,
508
+ "step": 830
509
+ },
510
+ {
511
+ "grad_norm": 1.0936087369918823,
512
+ "learning_rate": 9.318367465935142e-05,
513
+ "loss": 0.1306,
514
+ "step": 840
515
+ },
516
+ {
517
+ "grad_norm": 1.0435982942581177,
518
+ "learning_rate": 9.29738426991717e-05,
519
+ "loss": 0.1503,
520
+ "step": 850
521
+ },
522
+ {
523
+ "grad_norm": 0.9915235042572021,
524
+ "learning_rate": 9.276107353392774e-05,
525
+ "loss": 0.139,
526
+ "step": 860
527
+ },
528
+ {
529
+ "grad_norm": 1.1703134775161743,
530
+ "learning_rate": 9.254538170610938e-05,
531
+ "loss": 0.1373,
532
+ "step": 870
533
+ },
534
+ {
535
+ "grad_norm": 0.7841712236404419,
536
+ "learning_rate": 9.232678195796654e-05,
537
+ "loss": 0.1276,
538
+ "step": 880
539
+ },
540
+ {
541
+ "grad_norm": 1.0428593158721924,
542
+ "learning_rate": 9.210528923050164e-05,
543
+ "loss": 0.1221,
544
+ "step": 890
545
+ },
546
+ {
547
+ "grad_norm": 1.0131841897964478,
548
+ "learning_rate": 9.188091866244834e-05,
549
+ "loss": 0.1455,
550
+ "step": 900
551
+ },
552
+ {
553
+ "grad_norm": 1.151907205581665,
554
+ "learning_rate": 9.165368558923695e-05,
555
+ "loss": 0.1178,
556
+ "step": 910
557
+ },
558
+ {
559
+ "grad_norm": 1.0125540494918823,
560
+ "learning_rate": 9.142360554194618e-05,
561
+ "loss": 0.1403,
562
+ "step": 920
563
+ },
564
+ {
565
+ "grad_norm": 1.1062089204788208,
566
+ "learning_rate": 9.119069424624163e-05,
567
+ "loss": 0.125,
568
+ "step": 930
569
+ },
570
+ {
571
+ "grad_norm": 0.9944800138473511,
572
+ "learning_rate": 9.0954967621301e-05,
573
+ "loss": 0.117,
574
+ "step": 940
575
+ },
576
+ {
577
+ "grad_norm": 0.9723074436187744,
578
+ "learning_rate": 9.071644177872594e-05,
579
+ "loss": 0.14,
580
+ "step": 950
581
+ },
582
+ {
583
+ "grad_norm": 1.0174410343170166,
584
+ "learning_rate": 9.047513302144095e-05,
585
+ "loss": 0.1161,
586
+ "step": 960
587
+ },
588
+ {
589
+ "grad_norm": 0.8408202528953552,
590
+ "learning_rate": 9.023105784257906e-05,
591
+ "loss": 0.1185,
592
+ "step": 970
593
+ },
594
+ {
595
+ "grad_norm": 0.9733747839927673,
596
+ "learning_rate": 8.998423292435454e-05,
597
+ "loss": 0.1342,
598
+ "step": 980
599
+ },
600
+ {
601
+ "grad_norm": 0.9988760352134705,
602
+ "learning_rate": 8.973467513692265e-05,
603
+ "loss": 0.1289,
604
+ "step": 990
605
+ },
606
+ {
607
+ "grad_norm": 1.0810400247573853,
608
+ "learning_rate": 8.94824015372267e-05,
609
+ "loss": 0.1244,
610
+ "step": 1000
611
+ },
612
+ {
613
+ "grad_norm": 0.965705156326294,
614
+ "learning_rate": 8.922742936783207e-05,
615
+ "loss": 0.1122,
616
+ "step": 1010
617
+ },
618
+ {
619
+ "grad_norm": 1.09980309009552,
620
+ "learning_rate": 8.896977605574788e-05,
621
+ "loss": 0.1229,
622
+ "step": 1020
623
+ },
624
+ {
625
+ "grad_norm": 1.0067334175109863,
626
+ "learning_rate": 8.870945921123576e-05,
627
+ "loss": 0.1213,
628
+ "step": 1030
629
+ },
630
+ {
631
+ "grad_norm": 0.8215699791908264,
632
+ "learning_rate": 8.844649662660624e-05,
633
+ "loss": 0.1287,
634
+ "step": 1040
635
+ },
636
+ {
637
+ "grad_norm": 0.9577004909515381,
638
+ "learning_rate": 8.818090627500266e-05,
639
+ "loss": 0.1131,
640
+ "step": 1050
641
+ },
642
+ {
643
+ "grad_norm": 0.7915679216384888,
644
+ "learning_rate": 8.791270630917275e-05,
645
+ "loss": 0.1376,
646
+ "step": 1060
647
+ },
648
+ {
649
+ "grad_norm": 1.0239574909210205,
650
+ "learning_rate": 8.764191506022795e-05,
651
+ "loss": 0.1285,
652
+ "step": 1070
653
+ },
654
+ {
655
+ "grad_norm": 1.0196168422698975,
656
+ "learning_rate": 8.736855103639037e-05,
657
+ "loss": 0.1215,
658
+ "step": 1080
659
+ },
660
+ {
661
+ "grad_norm": 0.9371529817581177,
662
+ "learning_rate": 8.709263292172794e-05,
663
+ "loss": 0.0955,
664
+ "step": 1090
665
+ },
666
+ {
667
+ "grad_norm": 1.018921136856079,
668
+ "learning_rate": 8.681417957487729e-05,
669
+ "loss": 0.1159,
670
+ "step": 1100
671
+ },
672
+ {
673
+ "grad_norm": 0.9720669388771057,
674
+ "learning_rate": 8.653321002775478e-05,
675
+ "loss": 0.1245,
676
+ "step": 1110
677
+ },
678
+ {
679
+ "grad_norm": 0.9396262764930725,
680
+ "learning_rate": 8.624974348425574e-05,
681
+ "loss": 0.1169,
682
+ "step": 1120
683
+ },
684
+ {
685
+ "grad_norm": 0.8820628523826599,
686
+ "learning_rate": 8.596379931894188e-05,
687
+ "loss": 0.131,
688
+ "step": 1130
689
+ },
690
+ {
691
+ "grad_norm": 0.9641416072845459,
692
+ "learning_rate": 8.567539707571703e-05,
693
+ "loss": 0.1236,
694
+ "step": 1140
695
+ },
696
+ {
697
+ "grad_norm": 0.9010860323905945,
698
+ "learning_rate": 8.538455646649146e-05,
699
+ "loss": 0.1242,
700
+ "step": 1150
701
+ },
702
+ {
703
+ "grad_norm": 0.9010013937950134,
704
+ "learning_rate": 8.509129736983446e-05,
705
+ "loss": 0.1174,
706
+ "step": 1160
707
+ },
708
+ {
709
+ "grad_norm": 0.8815115690231323,
710
+ "learning_rate": 8.479563982961571e-05,
711
+ "loss": 0.1151,
712
+ "step": 1170
713
+ },
714
+ {
715
+ "grad_norm": 0.922867476940155,
716
+ "learning_rate": 8.449760405363539e-05,
717
+ "loss": 0.114,
718
+ "step": 1180
719
+ },
720
+ {
721
+ "grad_norm": 0.8651766777038574,
722
+ "learning_rate": 8.419721041224287e-05,
723
+ "loss": 0.1182,
724
+ "step": 1190
725
+ },
726
+ {
727
+ "grad_norm": 0.9760085344314575,
728
+ "learning_rate": 8.389447943694451e-05,
729
+ "loss": 0.1212,
730
+ "step": 1200
731
+ },
732
+ {
733
+ "grad_norm": 0.9760859608650208,
734
+ "learning_rate": 8.358943181900032e-05,
735
+ "loss": 0.1207,
736
+ "step": 1210
737
+ },
738
+ {
739
+ "grad_norm": 0.9782291054725647,
740
+ "learning_rate": 8.328208840800981e-05,
741
+ "loss": 0.1073,
742
+ "step": 1220
743
+ },
744
+ {
745
+ "grad_norm": 0.9667322039604187,
746
+ "learning_rate": 8.297247021048686e-05,
747
+ "loss": 0.1187,
748
+ "step": 1230
749
+ },
750
+ {
751
+ "grad_norm": 0.8478055596351624,
752
+ "learning_rate": 8.266059838842396e-05,
753
+ "loss": 0.1096,
754
+ "step": 1240
755
+ },
756
+ {
757
+ "grad_norm": 0.9771674275398254,
758
+ "learning_rate": 8.23464942578459e-05,
759
+ "loss": 0.1198,
760
+ "step": 1250
761
+ },
762
+ {
763
+ "grad_norm": 1.0028421878814697,
764
+ "learning_rate": 8.203017928735277e-05,
765
+ "loss": 0.1097,
766
+ "step": 1260
767
+ },
768
+ {
769
+ "grad_norm": 0.8741995096206665,
770
+ "learning_rate": 8.17116750966526e-05,
771
+ "loss": 0.1227,
772
+ "step": 1270
773
+ },
774
+ {
775
+ "grad_norm": 0.9604556560516357,
776
+ "learning_rate": 8.139100345508377e-05,
777
+ "loss": 0.1165,
778
+ "step": 1280
779
+ },
780
+ {
781
+ "grad_norm": 1.0160092115402222,
782
+ "learning_rate": 8.106818628012697e-05,
783
+ "loss": 0.1221,
784
+ "step": 1290
785
+ },
786
+ {
787
+ "grad_norm": 0.8472312092781067,
788
+ "learning_rate": 8.074324563590736e-05,
789
+ "loss": 0.1089,
790
+ "step": 1300
791
+ },
792
+ {
793
+ "grad_norm": 0.7550016641616821,
794
+ "learning_rate": 8.041620373168628e-05,
795
+ "loss": 0.1117,
796
+ "step": 1310
797
+ },
798
+ {
799
+ "grad_norm": 0.9404501914978027,
800
+ "learning_rate": 8.008708292034349e-05,
801
+ "loss": 0.1206,
802
+ "step": 1320
803
+ },
804
+ {
805
+ "grad_norm": 0.7967647314071655,
806
+ "learning_rate": 7.975590569684925e-05,
807
+ "loss": 0.1012,
808
+ "step": 1330
809
+ },
810
+ {
811
+ "grad_norm": 0.9530890583992004,
812
+ "learning_rate": 7.942269469672687e-05,
813
+ "loss": 0.1104,
814
+ "step": 1340
815
+ },
816
+ {
817
+ "grad_norm": 0.8970588445663452,
818
+ "learning_rate": 7.908747269450558e-05,
819
+ "loss": 0.0999,
820
+ "step": 1350
821
+ },
822
+ {
823
+ "grad_norm": 0.8660910129547119,
824
+ "learning_rate": 7.875026260216393e-05,
825
+ "loss": 0.1069,
826
+ "step": 1360
827
+ },
828
+ {
829
+ "grad_norm": 0.836333692073822,
830
+ "learning_rate": 7.841108746756382e-05,
831
+ "loss": 0.0878,
832
+ "step": 1370
833
+ },
834
+ {
835
+ "grad_norm": 1.0235843658447266,
836
+ "learning_rate": 7.806997047287516e-05,
837
+ "loss": 0.1074,
838
+ "step": 1380
839
+ },
840
+ {
841
+ "grad_norm": 0.9237326383590698,
842
+ "learning_rate": 7.772693493299138e-05,
843
+ "loss": 0.0956,
844
+ "step": 1390
845
+ },
846
+ {
847
+ "grad_norm": 0.9774718880653381,
848
+ "learning_rate": 7.7382004293936e-05,
849
+ "loss": 0.1126,
850
+ "step": 1400
851
+ },
852
+ {
853
+ "grad_norm": 0.94520503282547,
854
+ "learning_rate": 7.703520213126e-05,
855
+ "loss": 0.0951,
856
+ "step": 1410
857
+ },
858
+ {
859
+ "grad_norm": 0.8733212351799011,
860
+ "learning_rate": 7.66865521484305e-05,
861
+ "loss": 0.1121,
862
+ "step": 1420
863
+ },
864
+ {
865
+ "grad_norm": 0.8547873497009277,
866
+ "learning_rate": 7.633607817521074e-05,
867
+ "loss": 0.1034,
868
+ "step": 1430
869
+ },
870
+ {
871
+ "grad_norm": 0.701534628868103,
872
+ "learning_rate": 7.598380416603119e-05,
873
+ "loss": 0.106,
874
+ "step": 1440
875
+ },
876
+ {
877
+ "grad_norm": 0.9103834629058838,
878
+ "learning_rate": 7.562975419835247e-05,
879
+ "loss": 0.1101,
880
+ "step": 1450
881
+ },
882
+ {
883
+ "grad_norm": 0.8019031286239624,
884
+ "learning_rate": 7.527395247101956e-05,
885
+ "loss": 0.1018,
886
+ "step": 1460
887
+ },
888
+ {
889
+ "grad_norm": 0.9530763626098633,
890
+ "learning_rate": 7.491642330260789e-05,
891
+ "loss": 0.1055,
892
+ "step": 1470
893
+ },
894
+ {
895
+ "grad_norm": 0.8493635654449463,
896
+ "learning_rate": 7.45571911297612e-05,
897
+ "loss": 0.1098,
898
+ "step": 1480
899
+ },
900
+ {
901
+ "grad_norm": 0.9342841506004333,
902
+ "learning_rate": 7.419628050552131e-05,
903
+ "loss": 0.0957,
904
+ "step": 1490
905
+ },
906
+ {
907
+ "grad_norm": 0.8458704948425293,
908
+ "learning_rate": 7.383371609764999e-05,
909
+ "loss": 0.121,
910
+ "step": 1500
911
+ },
912
+ {
913
+ "grad_norm": 0.8073334097862244,
914
+ "learning_rate": 7.346952268694288e-05,
915
+ "loss": 0.1047,
916
+ "step": 1510
917
+ },
918
+ {
919
+ "grad_norm": 0.9626893401145935,
920
+ "learning_rate": 7.310372516553585e-05,
921
+ "loss": 0.1073,
922
+ "step": 1520
923
+ },
924
+ {
925
+ "grad_norm": 0.9793810844421387,
926
+ "learning_rate": 7.273634853520356e-05,
927
+ "loss": 0.1071,
928
+ "step": 1530
929
+ },
930
+ {
931
+ "grad_norm": 0.7965155243873596,
932
+ "learning_rate": 7.236741790565072e-05,
933
+ "loss": 0.1127,
934
+ "step": 1540
935
+ },
936
+ {
937
+ "grad_norm": 0.8083896040916443,
938
+ "learning_rate": 7.199695849279576e-05,
939
+ "loss": 0.1026,
940
+ "step": 1550
941
+ },
942
+ {
943
+ "grad_norm": 0.7700480818748474,
944
+ "learning_rate": 7.162499561704747e-05,
945
+ "loss": 0.1008,
946
+ "step": 1560
947
+ },
948
+ {
949
+ "grad_norm": 0.8370183706283569,
950
+ "learning_rate": 7.125155470157429e-05,
951
+ "loss": 0.0926,
952
+ "step": 1570
953
+ },
954
+ {
955
+ "grad_norm": 0.7990249395370483,
956
+ "learning_rate": 7.087666127056675e-05,
957
+ "loss": 0.0912,
958
+ "step": 1580
959
+ },
960
+ {
961
+ "grad_norm": 1.001415491104126,
962
+ "learning_rate": 7.050034094749286e-05,
963
+ "loss": 0.0995,
964
+ "step": 1590
965
+ },
966
+ {
967
+ "grad_norm": 0.7202123999595642,
968
+ "learning_rate": 7.012261945334683e-05,
969
+ "loss": 0.1017,
970
+ "step": 1600
971
+ },
972
+ {
973
+ "grad_norm": 0.7669859528541565,
974
+ "learning_rate": 6.974352260489103e-05,
975
+ "loss": 0.087,
976
+ "step": 1610
977
+ },
978
+ {
979
+ "grad_norm": 0.8397085070610046,
980
+ "learning_rate": 6.936307631289148e-05,
981
+ "loss": 0.0857,
982
+ "step": 1620
983
+ },
984
+ {
985
+ "grad_norm": 0.8605853915214539,
986
+ "learning_rate": 6.898130658034685e-05,
987
+ "loss": 0.0913,
988
+ "step": 1630
989
+ },
990
+ {
991
+ "grad_norm": 0.8747912049293518,
992
+ "learning_rate": 6.859823950071127e-05,
993
+ "loss": 0.1007,
994
+ "step": 1640
995
+ },
996
+ {
997
+ "grad_norm": 0.8252431750297546,
998
+ "learning_rate": 6.821390125611078e-05,
999
+ "loss": 0.1108,
1000
+ "step": 1650
1001
+ },
1002
+ {
1003
+ "grad_norm": 0.8187602758407593,
1004
+ "learning_rate": 6.782831811555385e-05,
1005
+ "loss": 0.0801,
1006
+ "step": 1660
1007
+ },
1008
+ {
1009
+ "grad_norm": 0.7218433022499084,
1010
+ "learning_rate": 6.744151643313597e-05,
1011
+ "loss": 0.097,
1012
+ "step": 1670
1013
+ },
1014
+ {
1015
+ "grad_norm": 0.8178585171699524,
1016
+ "learning_rate": 6.705352264623828e-05,
1017
+ "loss": 0.1048,
1018
+ "step": 1680
1019
+ },
1020
+ {
1021
+ "grad_norm": 0.7383034229278564,
1022
+ "learning_rate": 6.666436327372078e-05,
1023
+ "loss": 0.1006,
1024
+ "step": 1690
1025
+ },
1026
+ {
1027
+ "grad_norm": 0.9042794704437256,
1028
+ "learning_rate": 6.62740649141096e-05,
1029
+ "loss": 0.0956,
1030
+ "step": 1700
1031
+ },
1032
+ {
1033
+ "grad_norm": 0.823955237865448,
1034
+ "learning_rate": 6.588265424377919e-05,
1035
+ "loss": 0.0958,
1036
+ "step": 1710
1037
+ },
1038
+ {
1039
+ "grad_norm": 0.822413980960846,
1040
+ "learning_rate": 6.549015801512895e-05,
1041
+ "loss": 0.101,
1042
+ "step": 1720
1043
+ },
1044
+ {
1045
+ "grad_norm": 0.8048660159111023,
1046
+ "learning_rate": 6.509660305475468e-05,
1047
+ "loss": 0.0988,
1048
+ "step": 1730
1049
+ },
1050
+ {
1051
+ "grad_norm": 0.7963635325431824,
1052
+ "learning_rate": 6.47020162616152e-05,
1053
+ "loss": 0.0914,
1054
+ "step": 1740
1055
+ },
1056
+ {
1057
+ "grad_norm": 0.9308316111564636,
1058
+ "learning_rate": 6.430642460519365e-05,
1059
+ "loss": 0.0956,
1060
+ "step": 1750
1061
+ },
1062
+ {
1063
+ "grad_norm": 0.8854753971099854,
1064
+ "learning_rate": 6.390985512365426e-05,
1065
+ "loss": 0.101,
1066
+ "step": 1760
1067
+ },
1068
+ {
1069
+ "grad_norm": 0.8771095871925354,
1070
+ "learning_rate": 6.351233492199431e-05,
1071
+ "loss": 0.0972,
1072
+ "step": 1770
1073
+ },
1074
+ {
1075
+ "grad_norm": 0.775101900100708,
1076
+ "learning_rate": 6.311389117019155e-05,
1077
+ "loss": 0.0947,
1078
+ "step": 1780
1079
+ },
1080
+ {
1081
+ "grad_norm": 0.7772336602210999,
1082
+ "learning_rate": 6.271455110134713e-05,
1083
+ "loss": 0.1025,
1084
+ "step": 1790
1085
+ },
1086
+ {
1087
+ "grad_norm": 0.8436779975891113,
1088
+ "learning_rate": 6.231434200982428e-05,
1089
+ "loss": 0.0943,
1090
+ "step": 1800
1091
+ },
1092
+ {
1093
+ "grad_norm": 0.8183788061141968,
1094
+ "learning_rate": 6.191329124938285e-05,
1095
+ "loss": 0.0933,
1096
+ "step": 1810
1097
+ },
1098
+ {
1099
+ "grad_norm": 0.7852672934532166,
1100
+ "learning_rate": 6.15114262313095e-05,
1101
+ "loss": 0.0847,
1102
+ "step": 1820
1103
+ },
1104
+ {
1105
+ "grad_norm": 0.8776318430900574,
1106
+ "learning_rate": 6.110877442254444e-05,
1107
+ "loss": 0.0901,
1108
+ "step": 1830
1109
+ },
1110
+ {
1111
+ "grad_norm": 0.6940832734107971,
1112
+ "learning_rate": 6.0705363343803946e-05,
1113
+ "loss": 0.0866,
1114
+ "step": 1840
1115
+ },
1116
+ {
1117
+ "grad_norm": 0.6881867051124573,
1118
+ "learning_rate": 6.030122056769934e-05,
1119
+ "loss": 0.0913,
1120
+ "step": 1850
1121
+ },
1122
+ {
1123
+ "grad_norm": 0.8912978768348694,
1124
+ "learning_rate": 5.989637371685257e-05,
1125
+ "loss": 0.1051,
1126
+ "step": 1860
1127
+ },
1128
+ {
1129
+ "grad_norm": 0.6981568336486816,
1130
+ "learning_rate": 5.949085046200808e-05,
1131
+ "loss": 0.0984,
1132
+ "step": 1870
1133
+ },
1134
+ {
1135
+ "grad_norm": 0.8210259079933167,
1136
+ "learning_rate": 5.908467852014169e-05,
1137
+ "loss": 0.0831,
1138
+ "step": 1880
1139
+ },
1140
+ {
1141
+ "grad_norm": 0.7705551385879517,
1142
+ "learning_rate": 5.867788565256607e-05,
1143
+ "loss": 0.0748,
1144
+ "step": 1890
1145
+ },
1146
+ {
1147
+ "grad_norm": 0.7534878253936768,
1148
+ "learning_rate": 5.827049966303335e-05,
1149
+ "loss": 0.1015,
1150
+ "step": 1900
1151
+ },
1152
+ {
1153
+ "grad_norm": 0.8059409856796265,
1154
+ "learning_rate": 5.786254839583478e-05,
1155
+ "loss": 0.0944,
1156
+ "step": 1910
1157
+ },
1158
+ {
1159
+ "grad_norm": 0.9321911334991455,
1160
+ "learning_rate": 5.745405973389757e-05,
1161
+ "loss": 0.092,
1162
+ "step": 1920
1163
+ },
1164
+ {
1165
+ "grad_norm": 0.6626248359680176,
1166
+ "learning_rate": 5.7045061596879134e-05,
1167
+ "loss": 0.0904,
1168
+ "step": 1930
1169
+ },
1170
+ {
1171
+ "grad_norm": 0.7916935682296753,
1172
+ "learning_rate": 5.6635581939258855e-05,
1173
+ "loss": 0.0928,
1174
+ "step": 1940
1175
+ },
1176
+ {
1177
+ "grad_norm": 0.7326240539550781,
1178
+ "learning_rate": 5.622564874842742e-05,
1179
+ "loss": 0.0919,
1180
+ "step": 1950
1181
+ },
1182
+ {
1183
+ "grad_norm": 0.7140055894851685,
1184
+ "learning_rate": 5.5815290042773836e-05,
1185
+ "loss": 0.0756,
1186
+ "step": 1960
1187
+ },
1188
+ {
1189
+ "grad_norm": 0.7153354287147522,
1190
+ "learning_rate": 5.540453386977058e-05,
1191
+ "loss": 0.1021,
1192
+ "step": 1970
1193
+ },
1194
+ {
1195
+ "grad_norm": 0.6839438080787659,
1196
+ "learning_rate": 5.4993408304056425e-05,
1197
+ "loss": 0.1019,
1198
+ "step": 1980
1199
+ },
1200
+ {
1201
+ "grad_norm": 0.7680396437644958,
1202
+ "learning_rate": 5.458194144551768e-05,
1203
+ "loss": 0.0866,
1204
+ "step": 1990
1205
+ },
1206
+ {
1207
+ "grad_norm": 0.800545334815979,
1208
+ "learning_rate": 5.417016141736756e-05,
1209
+ "loss": 0.0779,
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": 2000,
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
+ }
a7-cabinet-close/b256/step-002000/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fce3fadf4ccc191593349fe1841cd110751b1046a4ed33e66c253e347b0b194c
3
+ size 8017
a7-cabinet-close/b256/step-002000/training_plan.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"task": "a7", "frames": 30664, "recipe": "full fine-tune (tune_llm, tune_visual, projector, diffusion, vlln), fixed steps", "per_gpu_batch": 32, "gradient_accumulation_steps": 1, "world_size": 8, "global_batch": 256, "max_steps": 4000, "save_steps": 2000, "epochs_equivalent": 33.39}
a7-cabinet-close/b256/step-002000/wandb_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"project": "finetune-gr00t-n1d7", "run_id": "production"}
a7-cabinet-close/b256/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)
a7-cabinet-close/b256/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)
a7-cabinet-close/b256/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)
a7-cabinet-close/b256/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")
a7-cabinet-close/b256/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)
a7-cabinet-close/b256/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)
a7-cabinet-close/b256/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)
a7-cabinet-close/b256/step-004000/code/hb_groot/hb_launch.py ADDED
@@ -0,0 +1,92 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ if os.environ.get("HB_PATCH_ONLY") != "1":
78
+ entry = Path(os.environ["GROOT_SRC"]) / "gr00t/experiment/launch_finetune.py"
79
+ sys.argv = [str(entry)] + sys.argv[1:]
80
+ try:
81
+ runpy.run_path(str(entry), run_name="__main__")
82
+ finally: # batch probe: report this rank's peak reserved memory (groot_chain.py compares it with MAX_FRAC)
83
+ _peak = os.environ.get("HB_PEAK_FILE")
84
+ if _peak and torch.cuda.is_available():
85
+ try:
86
+ _dev = torch.cuda.current_device()
87
+ with open(f"{_peak}-rank{os.environ.get('RANK', '0')}.json", "w") as _f:
88
+ json.dump({"rank": int(os.environ.get("RANK", "0")),
89
+ "peak_reserved_gb": torch.cuda.max_memory_reserved(_dev) / 1e9,
90
+ "total_gb": torch.cuda.get_device_properties(_dev).total_memory / 1e9}, _f)
91
+ except Exception: # noqa: BLE001 - never mask the training result
92
+ pass
a7-cabinet-close/b256/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
a7-cabinet-close/b256/step-004000/code/hb_groot/prepare_groot_data.py ADDED
@@ -0,0 +1,162 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "trolley": (P / "trolley-push-expert-v1-20261110/data/trolley_push", 100), # HF trolley-push-expert-v1 @ef9024a: fixed first 100 of 103
21
+ "b7": (P / "b7-black-clothes-hamper-v1-1bb002a/data", 100), # HF b7-black-clothes-hamper-expert-v1 @1bb002a, first 100 of 107
22
+ "a6": (P / "a6-door-pull-v1-cf3ee9b/data/door_pull", None), # HF door-pull-expert-v1 @cf3ee9b (A6)
23
+ "c6t": (P / "c6-test8-06c231d/data", None), # c6 8-episode adaptation test only
24
+ "c6": (P / "c6-bathroom-door-fetch-v1-3afba69/data", None), # HF c6-bathroom-door-fetch-expert-v1 @3afba69 (C6)
25
+ "a3": (P / "a3-a7-20261109/a3-fridge-open-expert-v1/780fcf1e4bd7db947a8c2d8f7b92e61f24f71b45", None),
26
+ "a7": (P / "a3-a7-20261109/a7-cabinet-close-expert-v1/1935a98e8f55d4d1d2d1615ddb97c2679713595c", None),
27
+ "a18": (P / "a18-h2-dataset-audit-8e956a10/dataset", None),
28
+ "a17": (P / "a17-unscrew-bottle-v1-407e6e2/data", None), # HF a17-unscrew-bottle-expert-v1 @407e6e2
29
+ "a14": (P / "a14-unscrew-bulb-v1-b4390fc/data", None), # HF a14-unscrew-bulb-expert-v1 @b4390fc (H2, 3 cameras)
30
+ }
31
+ CONFIG = {"a18": "h2w3_body31_hand44_h16.py", "a17": "h2w3_body31_hand44_h16.py", "a14": "h2w3_body31_hand44_h16.py"}
32
+ # model-input cameras per H2 task (user 2026-11-12 22:0x: EVERY H2 task uses head + chest + wrist; third_person never)
33
+ CAMS = {"a18": ("head", "chest", "wrist"), "a17": ("head", "chest", "wrist"), "a14": ("head", "chest", "wrist")}
34
+ H2_GROUPS = {"left_leg": (0, 6), "right_leg": (6, 12), "waist": (12, 15), "left_arm": (15, 22), "right_arm": (22, 29),
35
+ "head": (29, 31), "left_hand": (31, 53), "right_hand": (53, 75)}
36
+
37
+
38
+ def h2_modality(info, cams=("head", "chest")):
39
+ f = info["features"]
40
+ names = f["observation.state"]["names"]
41
+ assert len(names) == 75 and names[:31] == f["action.body_target"]["names"] and names[31:] == f["action.hands"]["names"]
42
+ assert f["observation.projected_gravity"]["shape"] == [3]
43
+ state = {g: {"original_key": "observation.state", "start": a, "end": b} for g, (a, b) in H2_GROUPS.items()}
44
+ state["projected_gravity"] = {"original_key": "observation.projected_gravity", "start": 0, "end": 3}
45
+ return {"state": state,
46
+ "action": {"body_target": {"original_key": "action.body_target", "start": 0, "end": 31},
47
+ "left_hand_target": {"original_key": "action.hands", "start": 0, "end": 22},
48
+ "right_hand_target": {"original_key": "action.hands", "start": 22, "end": 44}},
49
+ "video": {c: {"original_key": f"observation.images.{c}"} for c in cams},
50
+ "annotation": {"human.task_description": {"original_key": "task_index"}}}
51
+ BODY = ["left_hip_pitch_joint", "left_hip_roll_joint", "left_hip_yaw_joint", "left_knee_joint", "left_ankle_pitch_joint",
52
+ "left_ankle_roll_joint", "right_hip_pitch_joint", "right_hip_roll_joint", "right_hip_yaw_joint",
53
+ "right_knee_joint", "right_ankle_pitch_joint", "right_ankle_roll_joint", "waist_yaw_joint", "waist_roll_joint",
54
+ "waist_pitch_joint", "left_shoulder_pitch_joint", "left_shoulder_roll_joint", "left_shoulder_yaw_joint",
55
+ "left_elbow_joint", "left_wrist_roll_joint", "left_wrist_pitch_joint", "left_wrist_yaw_joint",
56
+ "right_shoulder_pitch_joint", "right_shoulder_roll_joint", "right_shoulder_yaw_joint", "right_elbow_joint",
57
+ "right_wrist_roll_joint", "right_wrist_pitch_joint", "right_wrist_yaw_joint"]
58
+ GROUPS = {"left_leg": (0, 6), "right_leg": (6, 12), "waist": (12, 15), "left_arm": (15, 22), "right_arm": (22, 29)}
59
+
60
+
61
+ def jl(path):
62
+ return [json.loads(x) for x in Path(path).read_text().splitlines() if x.strip()]
63
+
64
+
65
+ def wl(path, rows):
66
+ Path(path).write_text("".join(json.dumps(r) + "\n" for r in rows))
67
+
68
+
69
+ def build(task):
70
+ src, limit = TASKS[task]
71
+ info = json.loads((src / "meta/info.json").read_text())
72
+ names = info["features"]["observation.state"]["names"]
73
+ if task in CONFIG:
74
+ idx = list(range(31))
75
+ else:
76
+ idx = [names.index(n) for n in BODY]
77
+ for g, (a, b) in GROUPS.items():
78
+ assert idx[a:b] == list(range(idx[a], idx[a] + (b - a))), (task, g, idx[a:b])
79
+ for key, dim in (("action.motion_token", 64), ("observation.inspire_left_hand_state", 6),
80
+ ("observation.inspire_right_hand_state", 6), ("observation.projected_gravity", 3),
81
+ ("observation.images.ego_view", None), ("observation.images.chest", None)):
82
+ assert key in info["features"], (task, key)
83
+ if dim:
84
+ assert info["features"][key]["shape"] == [dim], (task, key)
85
+ episodes = jl(src / "meta/episodes.jsonl")
86
+ stats = jl(src / "meta/episodes_stats.jsonl")
87
+ n = len(episodes) if limit is None else limit
88
+ episodes, stats = episodes[:n], stats[:n]
89
+ assert [e["episode_index"] for e in episodes] == list(range(n)) == [s["episode_index"] for s in stats]
90
+ frames = sum(int(e["length"]) for e in episodes)
91
+ dst = OUT / task
92
+ if dst.exists():
93
+ raise SystemExit(f"{dst} exists")
94
+ (dst / "meta").mkdir(parents=True)
95
+ vkeys = [k for k, f in info["features"].items() if f.get("dtype") == "video"]
96
+ for e in episodes:
97
+ i = e["episode_index"]
98
+ chunk = i // int(info["chunks_size"])
99
+ rel = info["data_path"].format(episode_chunk=chunk, episode_index=i)
100
+ (dst / rel).parent.mkdir(parents=True, exist_ok=True)
101
+ os.symlink(src / rel, dst / rel)
102
+ for k in vkeys:
103
+ relv = info["video_path"].format(episode_chunk=chunk, video_key=k, episode_index=i)
104
+ (dst / relv).parent.mkdir(parents=True, exist_ok=True)
105
+ os.symlink(src / relv, dst / relv)
106
+ if task in CONFIG:
107
+ # stats.py computes stats for every low-dim feature in info.json and cannot handle e.g. observation.tactile
108
+ # [10,24,24]; keep only what the H2 modality uses (parquet columns stay untouched behind the symlinks).
109
+ keep = {"observation.state", "observation.projected_gravity", "action.body_target", "action.hands",
110
+ "timestamp", "frame_index", "episode_index", "index", "task_index"}
111
+ keep |= {f"observation.images.{c}" for c in CAMS[task]}
112
+ info["features"] = {k: v for k, v in info["features"].items() if k in keep}
113
+ vkeys = [k for k in vkeys if k in keep]
114
+ stats = [{**s, "stats": {k: v for k, v in s["stats"].items() if k in keep}} if "stats" in s else s
115
+ for s in stats]
116
+ info.update(total_episodes=n, total_frames=frames, total_videos=n * len(vkeys),
117
+ total_chunks=(n - 1) // int(info["chunks_size"]) + 1, splits={"train": f"0:{n}"})
118
+ (dst / "meta/info.json").write_text(json.dumps(info, indent=2))
119
+ wl(dst / "meta/episodes.jsonl", episodes)
120
+ wl(dst / "meta/episodes_stats.jsonl", stats)
121
+ (dst / "meta/tasks.jsonl").write_text((src / "meta/tasks.jsonl").read_text())
122
+ if task in CONFIG:
123
+ modality = h2_modality(info, CAMS[task])
124
+ else:
125
+ state = {g: {"original_key": "observation.state", "start": idx[a], "end": idx[a] + (b - a)}
126
+ for g, (a, b) in GROUPS.items()}
127
+ state.update(left_hand={"original_key": "observation.inspire_left_hand_state", "start": 0, "end": 6},
128
+ right_hand={"original_key": "observation.inspire_right_hand_state", "start": 0, "end": 6},
129
+ projected_gravity={"original_key": "observation.projected_gravity", "start": 0, "end": 3})
130
+ modality = {"state": state,
131
+ "action": {"motion_token": {"original_key": "action.motion_token", "start": 0, "end": 64},
132
+ "left_hand_joints": {"original_key": "observation.inspire_left_hand_state", "start": 0, "end": 6},
133
+ "right_hand_joints": {"original_key": "observation.inspire_right_hand_state", "start": 0, "end": 6}},
134
+ "video": {"ego_view": {"original_key": "observation.images.ego_view"},
135
+ "chest": {"original_key": "observation.images.chest"}},
136
+ "annotation": {"human.task_description": {"original_key": "task_index"}}}
137
+ (dst / "meta/modality.json").write_text(json.dumps(modality, indent=2))
138
+ py = sys.executable
139
+ src_g = G / "HumanWAM/third_party/Isaac-GR00T"
140
+ env = dict(os.environ, PYTHONPATH=f"{src_g}:{H}")
141
+ done = None
142
+ for tag in ("NEW_EMBODIMENT", "new_embodiment"):
143
+ r = subprocess.run([py, str(src_g / "gr00t/data/stats.py"), "--dataset-path", str(dst), "--embodiment-tag", tag,
144
+ "--modality-config-path", str(H / CONFIG.get(task, "g1_feedback76_h16.py"))], env=env, cwd=src_g,
145
+ capture_output=True, text=True)
146
+ if r.returncode == 0:
147
+ done = tag
148
+ break
149
+ last = r.stdout[-1500:] + r.stderr[-1500:]
150
+ if done is None:
151
+ raise SystemExit(f"stats failed for {task}:\n{last}")
152
+ assert (dst / "meta/stats.json").is_file()
153
+ receipt = dict(task=task, source=str(src), episodes=n, frames=frames, video_keys=vkeys,
154
+ state_dim_raw=len(names), body_index=idx, stats_tag=done,
155
+ modality_sha256=hashlib.sha256((dst / "meta/modality.json").read_bytes()).hexdigest(),
156
+ stats_sha256=hashlib.sha256((dst / "meta/stats.json").read_bytes()).hexdigest())
157
+ (dst / "hb-dataset-receipt.json").write_text(json.dumps(receipt, indent=2) + "\n")
158
+ print("DATASET_OK", json.dumps(receipt)[:400], flush=True)
159
+
160
+
161
+ for t in sys.argv[1:]:
162
+ build(t)
a7-cabinet-close/b256/step-004000/code/hb_groot/verify_groot_cpu.py ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 in ("a14", "a18", "a17"): # every H2 task: 3 cameras head + chest + wrist (user 2026-11-12)
20
+ import h2w3_body31_hand44_h16 as C # noqa: E402
21
+ cfg = C.h2w3_body31_hand44_h16
22
+ else:
23
+ import g1_feedback76_h16 as C # noqa: E402
24
+ cfg = C.feedback76_h16
25
+ loader = LeRobotEpisodeLoader(view, cfg)
26
+ df = loader[0]
27
+ cols = list(df.columns)
28
+ row = df.iloc[20]
29
+ out = {"episodes": len(loader), "columns": cols[:40], "rows_ep0": len(df)}
30
+ state_dims = {c: int(np.asarray(row[c]).size) for c in cols if c.startswith("state.")}
31
+ action_dims = {c: int(np.asarray(row[c]).size) for c in cols if c.startswith("action.")}
32
+ out.update(state_dims=state_dims, action_dims=action_dims, state_total=sum(state_dims.values()),
33
+ action_total=sum(action_dims.values()))
34
+ vids = [c for c in cols if c.startswith("video.")]
35
+ imgs = [np.asarray(row[c]) for c in vids]
36
+ out["video"] = {c: list(i.shape) for c, i in zip(vids, imgs)}
37
+ x = torch.stack([torch.from_numpy(i).permute(2, 0, 1).float().div(255) for i in imgs])[None]
38
+ rng = np.random.default_rng(0)
39
+ boosts = {}
40
+ for _ in range(60):
41
+ y, p = A.augment(x, rng)
42
+ boosts[p["boost"]] = boosts.get(p["boost"], 0) + 1
43
+ assert y.shape == x.shape and torch.isfinite(y).all() and float(y.min()) >= 0 and float(y.max()) <= 1
44
+ out["aug_boost_counts_60"] = boosts
45
+ out["mean_abs_change_last"] = round(float((y - x).abs().mean()), 4)
46
+ q = np.zeros((1, 29))
47
+ dq = A.perturb_joints(q, rng)
48
+ out["proprio_max_abs_deg"] = round(float(np.degrees(np.abs(dq).max())), 3)
49
+ out["patched"] = Gr00tN1d7Processor._get_vlm_inputs.__name__ == "_get_vlm_inputs" and \
50
+ Gr00tN1d7Processor.__call__.__module__ == "hb_launch"
51
+ print("CPU_VERIFY", json.dumps(out), flush=True)
a7-cabinet-close/b256/step-004000/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": 0.2,
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
+ }
a7-cabinet-close/b256/step-004000/dataset_meta/hb-dataset-receipt.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "a7",
3
+ "source": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/a3-a7-20261109/a7-cabinet-close-expert-v1/1935a98e8f55d4d1d2d1615ddb97c2679713595c",
4
+ "episodes": 100,
5
+ "frames": 30664,
6
+ "video_keys": [
7
+ "observation.images.ego_view",
8
+ "observation.images.chest"
9
+ ],
10
+ "state_dim_raw": 41,
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
+ ],
42
+ "stats_tag": "NEW_EMBODIMENT",
43
+ "modality_sha256": "aea724f383cf0aa4d74349163ac73ec7b1f6f30cf5a66efce0a12c50dfabf823",
44
+ "stats_sha256": "169e38827bb311d7b150aee2f59db9604c7bb219af7612fcaa63cb71e859d210"
45
+ }
a7-cabinet-close/b256/step-004000/dataset_meta/modality.json ADDED
@@ -0,0 +1,74 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "left_hand": {
29
+ "original_key": "observation.inspire_left_hand_state",
30
+ "start": 0,
31
+ "end": 6
32
+ },
33
+ "right_hand": {
34
+ "original_key": "observation.inspire_right_hand_state",
35
+ "start": 0,
36
+ "end": 6
37
+ },
38
+ "projected_gravity": {
39
+ "original_key": "observation.projected_gravity",
40
+ "start": 0,
41
+ "end": 3
42
+ }
43
+ },
44
+ "action": {
45
+ "motion_token": {
46
+ "original_key": "action.motion_token",
47
+ "start": 0,
48
+ "end": 64
49
+ },
50
+ "left_hand_joints": {
51
+ "original_key": "observation.inspire_left_hand_state",
52
+ "start": 0,
53
+ "end": 6
54
+ },
55
+ "right_hand_joints": {
56
+ "original_key": "observation.inspire_right_hand_state",
57
+ "start": 0,
58
+ "end": 6
59
+ }
60
+ },
61
+ "video": {
62
+ "ego_view": {
63
+ "original_key": "observation.images.ego_view"
64
+ },
65
+ "chest": {
66
+ "original_key": "observation.images.chest"
67
+ }
68
+ },
69
+ "annotation": {
70
+ "human.task_description": {
71
+ "original_key": "task_index"
72
+ }
73
+ }
74
+ }
a7-cabinet-close/b256/step-004000/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
+ }
a7-cabinet-close/b256/step-004000/experiment_cfg/conf.yaml ADDED
@@ -0,0 +1,234 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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: 0.2
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: 0.2
55
+ final_dropout: true
56
+ output_dim: 1024
57
+ interleave_self_attention: true
58
+ num_inference_timesteps: 4
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: 0.2
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/a7
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
+ - ego_view
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
+ - left_hand
99
+ - right_hand
100
+ - projected_gravity
101
+ sin_cos_embedding_keys: null
102
+ mean_std_embedding_keys: null
103
+ action_configs: null
104
+ action:
105
+ delta_indices:
106
+ - 0
107
+ - 1
108
+ - 2
109
+ - 3
110
+ - 4
111
+ - 5
112
+ - 6
113
+ - 7
114
+ - 8
115
+ - 9
116
+ - 10
117
+ - 11
118
+ - 12
119
+ - 13
120
+ - 14
121
+ - 15
122
+ modality_keys:
123
+ - motion_token
124
+ - left_hand_joints
125
+ - right_hand_joints
126
+ sin_cos_embedding_keys: null
127
+ mean_std_embedding_keys: null
128
+ action_configs:
129
+ - rep: ABSOLUTE
130
+ type: NON_EEF
131
+ format: DEFAULT
132
+ state_key: null
133
+ - rep: ABSOLUTE
134
+ type: NON_EEF
135
+ format: DEFAULT
136
+ state_key: null
137
+ - rep: ABSOLUTE
138
+ type: NON_EEF
139
+ format: DEFAULT
140
+ state_key: null
141
+ language:
142
+ delta_indices:
143
+ - 0
144
+ modality_keys:
145
+ - annotation.human.task_description
146
+ sin_cos_embedding_keys: null
147
+ mean_std_embedding_keys: null
148
+ action_configs: null
149
+ download_cache: false
150
+ shard_size: 1024
151
+ episode_sampling_rate: 0.1
152
+ num_shards_per_epoch: 100000
153
+ ds_weights_alpha: null
154
+ override_pretraining_statistics: true
155
+ mode: single_turn
156
+ random_chop: 0.0
157
+ mock_dataset_mode: false
158
+ shuffle: true
159
+ seed: 42
160
+ multiprocessing_context: fork
161
+ allow_padding: false
162
+ subsample_ratio: 1.0
163
+ loader: fast
164
+ fast_loader:
165
+ decode_threads: 2
166
+ ffmpeg_threads: 1
167
+ shuffle_buffer: 0
168
+ min_fill: 0
169
+ image_crop_size:
170
+ - 244
171
+ - 244
172
+ image_target_size:
173
+ - 224
174
+ - 224
175
+ training:
176
+ output_dir: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/groot-runs-ft4k/a7/production
177
+ experiment_name: null
178
+ max_steps: 4000
179
+ global_batch_size: 256
180
+ per_gpu_batch_size: null
181
+ gradient_accumulation_steps: 1
182
+ learning_rate: 0.0001
183
+ lr_scheduler_type: cosine
184
+ weight_decay: 1.0e-05
185
+ warmup_ratio: 0.05
186
+ warmup_steps: 0
187
+ max_grad_norm: 1.0
188
+ optim: adamw_torch
189
+ start_from_checkpoint: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/groot-n17-base
190
+ skip_weight_loading: false
191
+ tf32: true
192
+ fp16: false
193
+ bf16: true
194
+ eval_bf16: true
195
+ logging_steps: 10
196
+ save_steps: 2000
197
+ save_total_limit: 2
198
+ save_vl_model: false
199
+ save_only_model: false
200
+ resume_from_checkpoint: false
201
+ upload_checkpoints: false
202
+ upload_every: 1000
203
+ upload_last_n_checkpoints: 5
204
+ max_concurrent_uploads: 2
205
+ eval_strategy: 'no'
206
+ eval_steps: 500
207
+ eval_set_split_ratio: 0.1
208
+ eval_batch_size: 2
209
+ save_best_eval_metric_name: ''
210
+ save_best_eval_metric_greater_is_better: true
211
+ deepspeed_stage: 2
212
+ gradient_checkpointing: false
213
+ transformers_trust_remote_code: true
214
+ transformers_local_files_only: false
215
+ transformers_cache_dir: null
216
+ transformers_access_token: null
217
+ use_ddp: false
218
+ ddp_bucket_cap_mb: 100
219
+ num_gpus: 8
220
+ dataloader_num_workers: 8
221
+ remove_unused_columns: false
222
+ use_wandb: false
223
+ wandb_project: finetune-gr00t-n1d7
224
+ enable_profiling: false
225
+ max_retries: 3
226
+ assert_loss_less_than: null
227
+ add_rl_callback: false
228
+ enable_open_loop_eval: false
229
+ open_loop_eval_traj_ids:
230
+ - 0
231
+ open_loop_eval_steps_per_traj: 100
232
+ open_loop_eval_plot_indices: null
233
+ max_steps: 4000
234
+ save_steps: 2000
a7-cabinet-close/b256/step-004000/experiment_cfg/config.yaml ADDED
@@ -0,0 +1,232 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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: 0.2
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: 0.2
55
+ final_dropout: true
56
+ output_dim: 1024
57
+ interleave_self_attention: true
58
+ num_inference_timesteps: 4
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: 0.2
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/a7
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
+ - ego_view
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
+ - left_hand
99
+ - right_hand
100
+ - projected_gravity
101
+ sin_cos_embedding_keys: null
102
+ mean_std_embedding_keys: null
103
+ action_configs: null
104
+ action:
105
+ delta_indices:
106
+ - 0
107
+ - 1
108
+ - 2
109
+ - 3
110
+ - 4
111
+ - 5
112
+ - 6
113
+ - 7
114
+ - 8
115
+ - 9
116
+ - 10
117
+ - 11
118
+ - 12
119
+ - 13
120
+ - 14
121
+ - 15
122
+ modality_keys:
123
+ - motion_token
124
+ - left_hand_joints
125
+ - right_hand_joints
126
+ sin_cos_embedding_keys: null
127
+ mean_std_embedding_keys: null
128
+ action_configs:
129
+ - rep: absolute
130
+ type: non_eef
131
+ format: default
132
+ state_key: null
133
+ - rep: absolute
134
+ type: non_eef
135
+ format: default
136
+ state_key: null
137
+ - rep: absolute
138
+ type: non_eef
139
+ format: default
140
+ state_key: null
141
+ language:
142
+ delta_indices:
143
+ - 0
144
+ modality_keys:
145
+ - annotation.human.task_description
146
+ sin_cos_embedding_keys: null
147
+ mean_std_embedding_keys: null
148
+ action_configs: null
149
+ download_cache: false
150
+ shard_size: 1024
151
+ episode_sampling_rate: 0.1
152
+ num_shards_per_epoch: 100000
153
+ ds_weights_alpha: null
154
+ override_pretraining_statistics: true
155
+ mode: single_turn
156
+ random_chop: 0.0
157
+ mock_dataset_mode: false
158
+ shuffle: true
159
+ seed: 42
160
+ multiprocessing_context: fork
161
+ allow_padding: false
162
+ subsample_ratio: 1.0
163
+ loader: fast
164
+ fast_loader:
165
+ decode_threads: 2
166
+ ffmpeg_threads: 1
167
+ shuffle_buffer: 0
168
+ min_fill: 0
169
+ image_crop_size:
170
+ - 244
171
+ - 244
172
+ image_target_size:
173
+ - 224
174
+ - 224
175
+ training:
176
+ output_dir: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/groot-runs-ft4k/a7/production
177
+ experiment_name: null
178
+ max_steps: 4000
179
+ global_batch_size: 256
180
+ per_gpu_batch_size: null
181
+ gradient_accumulation_steps: 1
182
+ learning_rate: 0.0001
183
+ lr_scheduler_type: cosine
184
+ weight_decay: 1.0e-05
185
+ warmup_ratio: 0.05
186
+ warmup_steps: 0
187
+ max_grad_norm: 1.0
188
+ optim: adamw_torch
189
+ start_from_checkpoint: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/groot-n17-base
190
+ skip_weight_loading: false
191
+ tf32: true
192
+ fp16: false
193
+ bf16: true
194
+ eval_bf16: true
195
+ logging_steps: 10
196
+ save_steps: 2000
197
+ save_total_limit: 2
198
+ save_vl_model: false
199
+ save_only_model: false
200
+ resume_from_checkpoint: false
201
+ upload_checkpoints: false
202
+ upload_every: 1000
203
+ upload_last_n_checkpoints: 5
204
+ max_concurrent_uploads: 2
205
+ eval_strategy: 'no'
206
+ eval_steps: 500
207
+ eval_set_split_ratio: 0.1
208
+ eval_batch_size: 2
209
+ save_best_eval_metric_name: ''
210
+ save_best_eval_metric_greater_is_better: true
211
+ deepspeed_stage: 2
212
+ gradient_checkpointing: false
213
+ transformers_trust_remote_code: true
214
+ transformers_local_files_only: false
215
+ transformers_cache_dir: null
216
+ transformers_access_token: null
217
+ use_ddp: false
218
+ ddp_bucket_cap_mb: 100
219
+ num_gpus: 8
220
+ dataloader_num_workers: 8
221
+ remove_unused_columns: false
222
+ use_wandb: false
223
+ wandb_project: finetune-gr00t-n1d7
224
+ enable_profiling: false
225
+ max_retries: 3
226
+ assert_loss_less_than: null
227
+ add_rl_callback: false
228
+ enable_open_loop_eval: false
229
+ open_loop_eval_traj_ids:
230
+ - 0
231
+ open_loop_eval_steps_per_traj: 100
232
+ open_loop_eval_plot_indices: null
a7-cabinet-close/b256/step-004000/experiment_cfg/dataset_statistics.json ADDED
@@ -0,0 +1,883 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "new_embodiment": {
3
+ "state": {
4
+ "left_leg": {
5
+ "min": [
6
+ -0.9279626607894897,
7
+ -0.20401069521903992,
8
+ -0.5286000370979309,
9
+ -0.015552200376987457,
10
+ -0.4715784192085266,
11
+ -0.1929469108581543
12
+ ],
13
+ "max": [
14
+ 0.34397855401039124,
15
+ 0.31646499037742615,
16
+ 0.6314767599105835,
17
+ 1.2863916158676147,
18
+ 0.25699329376220703,
19
+ 0.18627917766571045
20
+ ],
21
+ "mean": [
22
+ -0.35561004281044006,
23
+ 0.037816599011421204,
24
+ 0.09485039860010147,
25
+ 0.5194756388664246,
26
+ -0.1980385035276413,
27
+ -0.006605133879929781
28
+ ],
29
+ "std": [
30
+ 0.28290456533432007,
31
+ 0.09233193844556808,
32
+ 0.19659870862960815,
33
+ 0.20531852543354034,
34
+ 0.08587498962879181,
35
+ 0.072576604783535
36
+ ],
37
+ "q01": [
38
+ -0.84334124147892,
39
+ -0.15387645348906517,
40
+ -0.3828186219930649,
41
+ 0.19774391919374465,
42
+ -0.4022857713699341,
43
+ -0.1577625685930252
44
+ ],
45
+ "q99": [
46
+ 0.19437677353620522,
47
+ 0.24248618379235268,
48
+ 0.45563414484262454,
49
+ 1.1226949226856227,
50
+ 0.00831889577209938,
51
+ 0.12483298838138578
52
+ ]
53
+ },
54
+ "right_leg": {
55
+ "min": [
56
+ -1.0070744752883911,
57
+ -0.3764217495918274,
58
+ -0.7747820019721985,
59
+ -0.07910795509815216,
60
+ -0.4874956011772156,
61
+ -0.21903759241104126
62
+ ],
63
+ "max": [
64
+ 0.6431071162223816,
65
+ 0.14184899628162384,
66
+ 0.18549136817455292,
67
+ 1.1398687362670898,
68
+ 0.33599478006362915,
69
+ 0.21929782629013062
70
+ ],
71
+ "mean": [
72
+ -0.3908240795135498,
73
+ -0.1153804287314415,
74
+ -0.2100031077861786,
75
+ 0.5305911898612976,
76
+ -0.18644081056118011,
77
+ -0.0369383879005909
78
+ ],
79
+ "std": [
80
+ 0.2948094308376312,
81
+ 0.09279008209705353,
82
+ 0.14734967052936554,
83
+ 0.20262227952480316,
84
+ 0.09397339075803757,
85
+ 0.07671832293272018
86
+ ],
87
+ "q01": [
88
+ -0.8900217366218567,
89
+ -0.3212410768866539,
90
+ -0.5801075077056885,
91
+ 0.16907327070832254,
92
+ -0.3691000935435295,
93
+ -0.17217106655240058
94
+ ],
95
+ "q99": [
96
+ 0.14047862306237197,
97
+ 0.06311478950083252,
98
+ 0.0607764326035973,
99
+ 0.9524855113029468,
100
+ 0.07608542144298523,
101
+ 0.15499507382512084
102
+ ]
103
+ },
104
+ "waist": {
105
+ "min": [
106
+ -0.07198242843151093,
107
+ -0.08106081932783127,
108
+ -0.09916144609451294
109
+ ],
110
+ "max": [
111
+ 0.19520391523838043,
112
+ 0.11672026664018631,
113
+ 0.11673737317323685
114
+ ],
115
+ "mean": [
116
+ 0.015539770945906639,
117
+ 0.013902479782700539,
118
+ 0.016449792310595512
119
+ ],
120
+ "std": [
121
+ 0.05012740567326546,
122
+ 0.035609565675258636,
123
+ 0.024248970672488213
124
+ ],
125
+ "q01": [
126
+ -0.050797894336283206,
127
+ -0.05091617964208126,
128
+ -0.048496404103934763
129
+ ],
130
+ "q99": [
131
+ 0.1525713104009628,
132
+ 0.08131846114993095,
133
+ 0.08361814886331549
134
+ ]
135
+ },
136
+ "left_arm": {
137
+ "min": [
138
+ 0.08239381015300751,
139
+ 0.015524790622293949,
140
+ -0.1121721938252449,
141
+ -0.4009973704814911,
142
+ -0.06022924929857254,
143
+ -0.1200670674443245,
144
+ -0.14843131601810455
145
+ ],
146
+ "max": [
147
+ 0.5195285677909851,
148
+ 0.22663770616054535,
149
+ 0.30364537239074707,
150
+ -0.12988236546516418,
151
+ 0.09782146662473679,
152
+ 0.08013442903757095,
153
+ 0.13720038533210754
154
+ ],
155
+ "mean": [
156
+ 0.20338064432144165,
157
+ 0.1432969570159912,
158
+ 0.03276205062866211,
159
+ -0.21513384580612183,
160
+ 0.004481739364564419,
161
+ 0.007482144050300121,
162
+ -0.02025328017771244
163
+ ],
164
+ "std": [
165
+ 0.03379302844405174,
166
+ 0.025766393169760704,
167
+ 0.035842087119817734,
168
+ 0.03536125645041466,
169
+ 0.012149316258728504,
170
+ 0.029369434341788292,
171
+ 0.037967316806316376
172
+ ],
173
+ "q01": [
174
+ 0.12020795546472073,
175
+ 0.07591435350477696,
176
+ -0.032209906913340094,
177
+ -0.3325797915458679,
178
+ -0.023349081706255673,
179
+ -0.07990536816418171,
180
+ -0.1029418321698904
181
+ ],
182
+ "q99": [
183
+ 0.32057825535535717,
184
+ 0.20605838119983666,
185
+ 0.15985203474760054,
186
+ -0.1695161056518555,
187
+ 0.04425690572708841,
188
+ 0.062308642081916304,
189
+ 0.0753356870263814
190
+ ]
191
+ },
192
+ "right_arm": {
193
+ "min": [
194
+ -1.307084083557129,
195
+ -0.42307642102241516,
196
+ -0.15481454133987427,
197
+ -0.590183675289154,
198
+ -0.09855499118566513,
199
+ -0.19881458580493927,
200
+ -0.2657955288887024
201
+ ],
202
+ "max": [
203
+ 0.7159279584884644,
204
+ 0.18037880957126617,
205
+ 0.3666689097881317,
206
+ 1.4362146854400635,
207
+ 0.14007999002933502,
208
+ 0.13200220465660095,
209
+ 0.18118087947368622
210
+ ],
211
+ "mean": [
212
+ -0.20058324933052063,
213
+ -0.07724551111459732,
214
+ 0.036659061908721924,
215
+ 0.21988210082054138,
216
+ 0.014881269074976444,
217
+ 0.009545939043164253,
218
+ 0.0062908814288675785
219
+ ],
220
+ "std": [
221
+ 0.47667089104652405,
222
+ 0.12800854444503784,
223
+ 0.08127856254577637,
224
+ 0.5201790928840637,
225
+ 0.031467556953430176,
226
+ 0.03061072714626789,
227
+ 0.08500789850950241
228
+ ],
229
+ "q01": [
230
+ -1.213955465555191,
231
+ -0.33947362542152404,
232
+ -0.08272466488182545,
233
+ -0.4007240226864815,
234
+ -0.06500550769269467,
235
+ -0.11041613429784775,
236
+ -0.21620410695672035
237
+ ],
238
+ "q99": [
239
+ 0.39352929234504636,
240
+ 0.16627973556518555,
241
+ 0.27784405529499046,
242
+ 1.3493560016155244,
243
+ 0.09599469408392904,
244
+ 0.07411817602813242,
245
+ 0.1364300513267517
246
+ ]
247
+ },
248
+ "left_hand": {
249
+ "min": [
250
+ 0.7841184139251709,
251
+ 0.9513298273086548,
252
+ 0.9574251770973206,
253
+ 0.9504732489585876,
254
+ 0.851092517375946,
255
+ 0.0
256
+ ],
257
+ "max": [
258
+ 1.0,
259
+ 1.0,
260
+ 1.0,
261
+ 1.0,
262
+ 1.0,
263
+ 0.30305954813957214
264
+ ],
265
+ "mean": [
266
+ 0.9994741678237915,
267
+ 0.9999215006828308,
268
+ 0.9999319911003113,
269
+ 0.9999361038208008,
270
+ 0.9997094869613647,
271
+ 0.0039056746754795313
272
+ ],
273
+ "std": [
274
+ 0.005472263786937151,
275
+ 0.0013332137605284705,
276
+ 0.0009734063060159317,
277
+ 0.000962596037425488,
278
+ 0.004112611990421562,
279
+ 0.028869736939668655
280
+ ],
281
+ "q01": [
282
+ 0.983048460483551,
283
+ 0.9994798117876053,
284
+ 0.9993437665700913,
285
+ 0.9995159441232682,
286
+ 0.9964931815862655,
287
+ 0.0
288
+ ],
289
+ "q99": [
290
+ 1.0,
291
+ 1.0,
292
+ 1.0,
293
+ 1.0,
294
+ 1.0,
295
+ 0.16704725056886674
296
+ ]
297
+ },
298
+ "right_hand": {
299
+ "min": [
300
+ 0.9825374484062195,
301
+ 0.9934524297714233,
302
+ 0.9735382199287415,
303
+ 0.9164777994155884,
304
+ 0.9381963014602661,
305
+ 0.0
306
+ ],
307
+ "max": [
308
+ 1.0,
309
+ 1.0,
310
+ 1.0,
311
+ 1.0,
312
+ 1.0,
313
+ 0.3030369281768799
314
+ ],
315
+ "mean": [
316
+ 0.9999874234199524,
317
+ 0.9999957084655762,
318
+ 0.9999594688415527,
319
+ 0.9998860359191895,
320
+ 0.9993945956230164,
321
+ 0.004138351418077946
322
+ ],
323
+ "std": [
324
+ 0.00026499829262562324,
325
+ 7.441091306000585e-05,
326
+ 0.0006803626310974737,
327
+ 0.001620433293278436,
328
+ 0.002877759747200986,
329
+ 0.0288485586643219
330
+ ],
331
+ "q01": [
332
+ 0.9998595207929611,
333
+ 0.9998556685447693,
334
+ 0.99947749376297,
335
+ 0.9987725448608399,
336
+ 0.9853247946500778,
337
+ 0.0
338
+ ],
339
+ "q99": [
340
+ 1.0,
341
+ 1.0,
342
+ 1.0,
343
+ 1.0,
344
+ 1.0,
345
+ 0.16710025072097778
346
+ ]
347
+ },
348
+ "projected_gravity": {
349
+ "min": [
350
+ -0.29140183329582214,
351
+ -0.18303944170475006,
352
+ -0.9999998807907104
353
+ ],
354
+ "max": [
355
+ 0.4053841233253479,
356
+ 0.10588564723730087,
357
+ -0.9137668013572693
358
+ ],
359
+ "mean": [
360
+ 0.08264563977718353,
361
+ -0.01804521679878235,
362
+ -0.986268162727356
363
+ ],
364
+ "std": [
365
+ 0.13568265736103058,
366
+ 0.0389891155064106,
367
+ 0.015453053638340727
368
+ ],
369
+ "q01": [
370
+ -0.13602852791547776,
371
+ -0.12972312435507774,
372
+ -0.999962568283081
373
+ ],
374
+ "q99": [
375
+ 0.325051891207695,
376
+ 0.0649579370766878,
377
+ -0.9452181661128998
378
+ ]
379
+ }
380
+ },
381
+ "action": {
382
+ "motion_token": {
383
+ "min": [
384
+ -0.0625,
385
+ -0.25,
386
+ -0.5625,
387
+ -0.375,
388
+ -0.3125,
389
+ -0.0625,
390
+ -0.0625,
391
+ -0.3125,
392
+ -0.125,
393
+ -0.3125,
394
+ -0.375,
395
+ -0.3125,
396
+ -0.375,
397
+ -0.125,
398
+ -0.0625,
399
+ -0.4375,
400
+ -0.1875,
401
+ -0.3125,
402
+ -0.4375,
403
+ -0.25,
404
+ -0.3125,
405
+ -0.4375,
406
+ -0.1875,
407
+ -0.1875,
408
+ 0.0,
409
+ -0.1875,
410
+ -0.3125,
411
+ -0.1875,
412
+ -0.1875,
413
+ -0.25,
414
+ -0.125,
415
+ -0.3125,
416
+ -0.3125,
417
+ -0.0625,
418
+ -0.1875,
419
+ -0.25,
420
+ -0.1875,
421
+ -0.375,
422
+ 0.0,
423
+ -0.1875,
424
+ -0.25,
425
+ -0.3125,
426
+ -0.125,
427
+ -0.25,
428
+ -0.125,
429
+ -0.3125,
430
+ -0.4375,
431
+ -0.5625,
432
+ -0.1875,
433
+ -0.5,
434
+ -0.25,
435
+ -0.1875,
436
+ -0.1875,
437
+ -0.3125,
438
+ -0.125,
439
+ -0.1875,
440
+ -0.4375,
441
+ -0.125,
442
+ -0.125,
443
+ -0.3125,
444
+ -0.25,
445
+ -0.1875,
446
+ -0.0625,
447
+ -0.25
448
+ ],
449
+ "max": [
450
+ 0.25,
451
+ 0.25,
452
+ 0.0625,
453
+ 0.3125,
454
+ 0.25,
455
+ 0.3125,
456
+ 0.3125,
457
+ 0.1875,
458
+ 0.1875,
459
+ -0.0625,
460
+ 0.125,
461
+ 0.125,
462
+ 0.125,
463
+ 0.25,
464
+ 0.5,
465
+ 0.25,
466
+ 0.1875,
467
+ 0.25,
468
+ 0.125,
469
+ 0.25,
470
+ 0.25,
471
+ 0.125,
472
+ 0.3125,
473
+ 0.4375,
474
+ 0.25,
475
+ 0.25,
476
+ 0.0,
477
+ 0.3125,
478
+ 0.375,
479
+ 0.1875,
480
+ 0.375,
481
+ 0.125,
482
+ 0.25,
483
+ 0.375,
484
+ 0.125,
485
+ 0.1875,
486
+ 0.4375,
487
+ 0.1875,
488
+ 0.4375,
489
+ 0.25,
490
+ 0.25,
491
+ 0.125,
492
+ 0.25,
493
+ 0.125,
494
+ 0.375,
495
+ 0.0625,
496
+ 0.1875,
497
+ 0.0625,
498
+ 0.3125,
499
+ 0.0625,
500
+ 0.375,
501
+ 0.25,
502
+ 0.5,
503
+ 0.0625,
504
+ 0.25,
505
+ 0.5,
506
+ 0.0625,
507
+ 0.3125,
508
+ 0.25,
509
+ 0.1875,
510
+ 0.25,
511
+ 0.1875,
512
+ 0.375,
513
+ 0.1875
514
+ ],
515
+ "mean": [
516
+ 0.1088450625538826,
517
+ 0.005287144333124161,
518
+ -0.19314587116241455,
519
+ 0.032281339168548584,
520
+ -0.026981966570019722,
521
+ 0.10573270171880722,
522
+ 0.15308871865272522,
523
+ -0.03123573213815689,
524
+ 0.004007142037153244,
525
+ -0.19649875164031982,
526
+ -0.06992931663990021,
527
+ -0.07310078293085098,
528
+ -0.10089192539453506,
529
+ 0.10953602194786072,
530
+ 0.1817685216665268,
531
+ -0.08574794232845306,
532
+ 0.05256571248173714,
533
+ -0.00392561312764883,
534
+ -0.16298019886016846,
535
+ 0.02755674347281456,
536
+ 0.009571484290063381,
537
+ -0.08721546083688736,
538
+ 0.03227726370096207,
539
+ 0.14658068120479584,
540
+ 0.09666057676076889,
541
+ 0.033968985080718994,
542
+ -0.14083494246006012,
543
+ 0.05588189885020256,
544
+ 0.10194160789251328,
545
+ -0.02117711305618286,
546
+ 0.1287054866552353,
547
+ -0.08698106557130814,
548
+ -0.0035730008967220783,
549
+ 0.15965178608894348,
550
+ -0.01520716492086649,
551
+ 0.007881799712777138,
552
+ 0.041449256241321564,
553
+ -0.08795329928398132,
554
+ 0.1488349586725235,
555
+ 0.010747537948191166,
556
+ 0.022936098277568817,
557
+ -0.09044400602579117,
558
+ 0.05614686757326126,
559
+ -0.1165250763297081,
560
+ 0.08106615394353867,
561
+ -0.10055765509605408,
562
+ -0.09149572998285294,
563
+ -0.21917599439620972,
564
+ 0.00674651050940156,
565
+ -0.15472948551177979,
566
+ 0.09517268091440201,
567
+ 0.0023908328730612993,
568
+ 0.08028755336999893,
569
+ -0.11982903629541397,
570
+ 0.09656070917844772,
571
+ 0.1366056352853775,
572
+ -0.13778983056545258,
573
+ 0.07862232625484467,
574
+ 0.08546259254217148,
575
+ -0.0442986898124218,
576
+ -0.052129533141851425,
577
+ 0.0008152882801368833,
578
+ 0.157947838306427,
579
+ -0.031631145626306534
580
+ ],
581
+ "std": [
582
+ 0.0668804869055748,
583
+ 0.12095443159341812,
584
+ 0.12842515110969543,
585
+ 0.14495190978050232,
586
+ 0.1207330971956253,
587
+ 0.06679413467645645,
588
+ 0.08728994429111481,
589
+ 0.12497754395008087,
590
+ 0.07046829164028168,
591
+ 0.04337522014975548,
592
+ 0.07340089976787567,
593
+ 0.08057640492916107,
594
+ 0.10612399876117706,
595
+ 0.05006479471921921,
596
+ 0.10516314208507538,
597
+ 0.14025066792964935,
598
+ 0.08458932489156723,
599
+ 0.09508704394102097,
600
+ 0.10922928899526596,
601
+ 0.11999057978391647,
602
+ 0.11393564939498901,
603
+ 0.08346482366323471,
604
+ 0.09646101295948029,
605
+ 0.11847557872533798,
606
+ 0.041381414979696274,
607
+ 0.11736250668764114,
608
+ 0.039891693741083145,
609
+ 0.10182952135801315,
610
+ 0.08461461961269379,
611
+ 0.09703998267650604,
612
+ 0.08291352540254593,
613
+ 0.07950283586978912,
614
+ 0.12207935005426407,
615
+ 0.07671988010406494,
616
+ 0.06723712384700775,
617
+ 0.0650196447968483,
618
+ 0.10957574844360352,
619
+ 0.12713782489299774,
620
+ 0.07227926701307297,
621
+ 0.11771422624588013,
622
+ 0.09277141839265823,
623
+ 0.06503032147884369,
624
+ 0.06486136466264725,
625
+ 0.0484318882226944,
626
+ 0.10573384165763855,
627
+ 0.05409308895468712,
628
+ 0.1513296663761139,
629
+ 0.13459795713424683,
630
+ 0.09658853709697723,
631
+ 0.11133205890655518,
632
+ 0.09778749197721481,
633
+ 0.12356267124414444,
634
+ 0.17755742371082306,
635
+ 0.04896029829978943,
636
+ 0.07157403975725174,
637
+ 0.1577388197183609,
638
+ 0.10338503122329712,
639
+ 0.11795450747013092,
640
+ 0.060382284224033356,
641
+ 0.11109492182731628,
642
+ 0.1049015000462532,
643
+ 0.06797655671834946,
644
+ 0.07634958624839783,
645
+ 0.07761601358652115
646
+ ],
647
+ "q01": [
648
+ 0.0,
649
+ -0.25,
650
+ -0.4375,
651
+ -0.25,
652
+ -0.25,
653
+ -0.0625,
654
+ -0.0625,
655
+ -0.25,
656
+ -0.125,
657
+ -0.25,
658
+ -0.25,
659
+ -0.25,
660
+ -0.3125,
661
+ -0.0625,
662
+ 0.0,
663
+ -0.375,
664
+ -0.125,
665
+ -0.1875,
666
+ -0.375,
667
+ -0.1875,
668
+ -0.25,
669
+ -0.3125,
670
+ -0.125,
671
+ -0.125,
672
+ 0.0,
673
+ -0.1875,
674
+ -0.25,
675
+ -0.125,
676
+ -0.125,
677
+ -0.1875,
678
+ -0.0625,
679
+ -0.25,
680
+ -0.25,
681
+ 0.0,
682
+ -0.125,
683
+ -0.1875,
684
+ -0.125,
685
+ -0.375,
686
+ 0.0,
687
+ -0.1875,
688
+ -0.1875,
689
+ -0.25,
690
+ -0.0625,
691
+ -0.1875,
692
+ -0.0625,
693
+ -0.25,
694
+ -0.375,
695
+ -0.5,
696
+ -0.1875,
697
+ -0.4375,
698
+ -0.1875,
699
+ -0.1875,
700
+ -0.1875,
701
+ -0.25,
702
+ -0.0625,
703
+ -0.125,
704
+ -0.375,
705
+ -0.125,
706
+ -0.0625,
707
+ -0.25,
708
+ -0.1875,
709
+ -0.125,
710
+ -0.0625,
711
+ -0.1875
712
+ ],
713
+ "q99": [
714
+ 0.25,
715
+ 0.25,
716
+ 0.0,
717
+ 0.3125,
718
+ 0.1875,
719
+ 0.25,
720
+ 0.3125,
721
+ 0.1875,
722
+ 0.125,
723
+ -0.125,
724
+ 0.0625,
725
+ 0.0625,
726
+ 0.0625,
727
+ 0.1875,
728
+ 0.4375,
729
+ 0.1875,
730
+ 0.1875,
731
+ 0.1875,
732
+ 0.0625,
733
+ 0.25,
734
+ 0.1875,
735
+ 0.0625,
736
+ 0.25,
737
+ 0.375,
738
+ 0.1875,
739
+ 0.25,
740
+ -0.0625,
741
+ 0.25,
742
+ 0.25,
743
+ 0.1875,
744
+ 0.3125,
745
+ 0.0625,
746
+ 0.25,
747
+ 0.3125,
748
+ 0.125,
749
+ 0.125,
750
+ 0.375,
751
+ 0.125,
752
+ 0.3125,
753
+ 0.25,
754
+ 0.1875,
755
+ 0.0625,
756
+ 0.1875,
757
+ 0.0,
758
+ 0.3125,
759
+ 0.0,
760
+ 0.125,
761
+ 0.0,
762
+ 0.1875,
763
+ 0.0625,
764
+ 0.3125,
765
+ 0.1875,
766
+ 0.4375,
767
+ 0.0,
768
+ 0.1875,
769
+ 0.4375,
770
+ 0.0,
771
+ 0.25,
772
+ 0.1875,
773
+ 0.1875,
774
+ 0.1875,
775
+ 0.125,
776
+ 0.3125,
777
+ 0.125
778
+ ]
779
+ },
780
+ "left_hand_joints": {
781
+ "min": [
782
+ 0.7841184139251709,
783
+ 0.9513298273086548,
784
+ 0.9574251770973206,
785
+ 0.9504732489585876,
786
+ 0.851092517375946,
787
+ 0.0
788
+ ],
789
+ "max": [
790
+ 1.0,
791
+ 1.0,
792
+ 1.0,
793
+ 1.0,
794
+ 1.0,
795
+ 0.30305954813957214
796
+ ],
797
+ "mean": [
798
+ 0.9994741678237915,
799
+ 0.9999215006828308,
800
+ 0.9999319911003113,
801
+ 0.9999361038208008,
802
+ 0.9997094869613647,
803
+ 0.0039056746754795313
804
+ ],
805
+ "std": [
806
+ 0.005472263786937151,
807
+ 0.0013332137605284705,
808
+ 0.0009734063060159317,
809
+ 0.000962596037425488,
810
+ 0.004112611990421562,
811
+ 0.028869736939668655
812
+ ],
813
+ "q01": [
814
+ 0.983048460483551,
815
+ 0.9994798117876053,
816
+ 0.9993437665700913,
817
+ 0.9995159441232682,
818
+ 0.9964931815862655,
819
+ 0.0
820
+ ],
821
+ "q99": [
822
+ 1.0,
823
+ 1.0,
824
+ 1.0,
825
+ 1.0,
826
+ 1.0,
827
+ 0.16704725056886674
828
+ ]
829
+ },
830
+ "right_hand_joints": {
831
+ "min": [
832
+ 0.9825374484062195,
833
+ 0.9934524297714233,
834
+ 0.9735382199287415,
835
+ 0.9164777994155884,
836
+ 0.9381963014602661,
837
+ 0.0
838
+ ],
839
+ "max": [
840
+ 1.0,
841
+ 1.0,
842
+ 1.0,
843
+ 1.0,
844
+ 1.0,
845
+ 0.3030369281768799
846
+ ],
847
+ "mean": [
848
+ 0.9999874234199524,
849
+ 0.9999957084655762,
850
+ 0.9999594688415527,
851
+ 0.9998860359191895,
852
+ 0.9993945956230164,
853
+ 0.004138351418077946
854
+ ],
855
+ "std": [
856
+ 0.00026499829262562324,
857
+ 7.441091306000585e-05,
858
+ 0.0006803626310974737,
859
+ 0.001620433293278436,
860
+ 0.002877759747200986,
861
+ 0.0288485586643219
862
+ ],
863
+ "q01": [
864
+ 0.9998595207929611,
865
+ 0.9998556685447693,
866
+ 0.99947749376297,
867
+ 0.9987725448608399,
868
+ 0.9853247946500778,
869
+ 0.0
870
+ ],
871
+ "q99": [
872
+ 1.0,
873
+ 1.0,
874
+ 1.0,
875
+ 1.0,
876
+ 1.0,
877
+ 0.16710025072097778
878
+ ]
879
+ }
880
+ },
881
+ "relative_action": {}
882
+ }
883
+ }
a7-cabinet-close/b256/step-004000/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": 0.2,
55
+ "exclude_state": false,
56
+ "use_mean_std": false,
57
+ "max_num_embodiments": 32
58
+ }