BIue-Jay commited on
Commit
40bb883
·
verified ·
1 Parent(s): b2043de

Update README and adapter code

Browse files
README.md CHANGED
@@ -6,19 +6,21 @@ tags:
6
  ---
7
  # Psi-Zero — Humanoid Benchmark
8
 
9
- Psi-Zero (Psi0) checkpoints fine-tuned on Humanoid Benchmark tasks (2026-11-10 training round).
10
 
11
  | Folder | Task | Robot |
12
  |---|---|---|
13
  | `b6-pedal-bin/` | Pick up a chip barrel, step on the pedal bin, drop it in | Unitree G1 |
14
  | `a3-fridge-open/` | Open the fridge door | Unitree G1 |
15
  | `a7-cabinet-close/` | Close the cabinet door | Unitree G1 |
16
- | `a18-pour-water/` | Pour water from a pitcher into a glass | H2 |
 
17
 
18
- Each task folder has `epoch-010/` and `epoch-020/` (100 demonstrations per task, 20 epochs, saved at epoch 10 and 20).
19
- Every epoch folder is self-contained: weights, normalization statistics, training config and `protocol.json`
20
- (dimensions, input/output layout, step counts, file list). Optimizer / resume state is not included.
21
- `code/` holds the adapter code used for training and inference.
22
 
23
- Training used the Humanoid Benchmark online data augmentation spec v2 (photometric, geometric, sensor and lighting
24
- augmentation plus proprioception noise/delay/drop), with the model's own image augmentation turned off.
 
 
6
  ---
7
  # Psi-Zero — Humanoid Benchmark
8
 
9
+ Psi-Zero (Psi0 action expert, flow matching) checkpoints trained on Humanoid Benchmark tasks (100 demonstrations per task). 4 nodes x 8 GPUs, 20 epochs, saved at epoch 10 and 20.
10
 
11
  | Folder | Task | Robot |
12
  |---|---|---|
13
  | `b6-pedal-bin/` | Pick up a chip barrel, step on the pedal bin, drop it in | Unitree G1 |
14
  | `a3-fridge-open/` | Open the fridge door | Unitree G1 |
15
  | `a7-cabinet-close/` | Close the cabinet door | Unitree G1 |
16
+ | `a18-pour-water/` | Pour water from a pitcher into a glass | H2 (Sharpa hands) |
17
+ | `a10-trolley-push/` | Push a trolley | Unitree G1 |
18
 
19
+ Each task folder has one folder per published checkpoint (`epoch-XXX/`). Every epoch folder is self-contained for
20
+ evaluation and deployment (weights, configuration, normalization statistics, `protocol.json`) and also carries the
21
+ simple training records kept for reproducibility (training plan, logs, metrics). Optimizer / scheduler / RNG /
22
+ DeepSpeed resume state is not included. `code/` holds the adapter code used for training and inference.
23
 
24
+ Model inputs are identical across all benchmark models: G1 state 44-D (29 body joint angles + 6 + 6 measured hand
25
+ values + 3 projected gravity), H2 state 78-D (75 joint angles + 3 projected gravity), no joint velocities.
26
+ Training used the Humanoid Benchmark online data augmentation spec v2 with the model's own image augmentation off.
code/hb_psi0/hb_args.py CHANGED
@@ -21,7 +21,7 @@ from pathlib import Path
21
 
22
  H = Path(__file__).resolve().parent
23
  G = Path(os.environ.get("G", H.parent))
24
- TASKS = ("b6", "basket", "a3", "a7", "a18")
25
  WORLD = int(os.environ.get("HB_WORLD", "32")) # 4 nodes x 8 GPUs
26
  EPOCHS, SAVE_EPOCH, WARMUP_FRAC = 20, 10, 0.025 # warmup: native 1000/40000 = 2.5 %
27
  HORIZON = 30
 
21
 
22
  H = Path(__file__).resolve().parent
23
  G = Path(os.environ.get("G", H.parent))
24
+ TASKS = ("b6", "basket", "a3", "a7", "a18", "trolley", "a17", "b7")
25
  WORLD = int(os.environ.get("HB_WORLD", "32")) # 4 nodes x 8 GPUs
26
  EPOCHS, SAVE_EPOCH, WARMUP_FRAC = 20, 10, 0.025 # warmup: native 1000/40000 = 2.5 %
27
  HORIZON = 30
code/hb_psi0/prepare_data.py CHANGED
@@ -45,10 +45,13 @@ P = G.parent
45
  OUT = Path(os.environ.get("HB_DATA_ROOT", G / "psi0-data"))
46
  TASKS = { # task -> (source dataset, episode limit, robot layout)
47
  "b6": (P / "b6-v5t7-ready-start-60ep-20261109/data/deliveries/g1/b6-v5t7-ready-start-20261109", 100, "g1"),
48
- "basket": (P / "g1-basket-transfer-ed77719-40epochs/dataset", None, "g1"),
 
 
49
  "a3": (P / "a3-a7-20261109/a3-fridge-open-expert-v1/780fcf1e4bd7db947a8c2d8f7b92e61f24f71b45", None, "g1"),
50
  "a7": (P / "a3-a7-20261109/a7-cabinet-close-expert-v1/1935a98e8f55d4d1d2d1615ddb97c2679713595c", None, "g1"),
51
  "a18": (P / "a18-h2-dataset-audit-8e956a10/dataset", None, "h2"),
 
52
  }
53
  BODY = ["left_hip_pitch_joint", "left_hip_roll_joint", "left_hip_yaw_joint", "left_knee_joint", "left_ankle_pitch_joint",
54
  "left_ankle_roll_joint", "right_hip_pitch_joint", "right_hip_roll_joint", "right_hip_yaw_joint",
 
45
  OUT = Path(os.environ.get("HB_DATA_ROOT", G / "psi0-data"))
46
  TASKS = { # task -> (source dataset, episode limit, robot layout)
47
  "b6": (P / "b6-v5t7-ready-start-60ep-20261109/data/deliveries/g1/b6-v5t7-ready-start-20261109", 100, "g1"),
48
+ "basket": (P / "basket-current-dynamics-100-v21-b0ff522/data", None, "g1"), # HF basket-transfer-expert-v1 @current-dynamics-100-v21 b0ff522 (old ed77719 retired)
49
+ "trolley": (P / "trolley-push-expert-v1-20261110/data/trolley_push", 100, "g1"), # HF trolley-push-expert-v1 @ef9024a: fixed first 100 of 103
50
+ "b7": (P / "b7-black-clothes-hamper-v1-1bb002a/data", 100, "g1"), # HF b7-black-clothes-hamper-expert-v1 @1bb002a, first 100 of 107
51
  "a3": (P / "a3-a7-20261109/a3-fridge-open-expert-v1/780fcf1e4bd7db947a8c2d8f7b92e61f24f71b45", None, "g1"),
52
  "a7": (P / "a3-a7-20261109/a7-cabinet-close-expert-v1/1935a98e8f55d4d1d2d1615ddb97c2679713595c", None, "g1"),
53
  "a18": (P / "a18-h2-dataset-audit-8e956a10/dataset", None, "h2"),
54
+ "a17": (P / "a17-unscrew-bottle-v1-407e6e2/data", None, "h2"), # HF a17-unscrew-bottle-expert-v1 @407e6e2
55
  }
56
  BODY = ["left_hip_pitch_joint", "left_hip_roll_joint", "left_hip_yaw_joint", "left_knee_joint", "left_ankle_pitch_joint",
57
  "left_ankle_roll_joint", "right_hip_pitch_joint", "right_hip_roll_joint", "right_hip_yaw_joint",