cloudwalk-research/psi0-g1-sneaker-325ep-v3

Psi0 G1 Sneaker-in-Box — 325 episodes (v3 training)

325 teleoperated episodes of a Unitree G1 humanoid (with Inspire RH56DFTP hands) on the "Pick up the sneaker and place it inside the box." task. Extends the psi0-g1-sneaker-205ep-v2-source canonical with 60 lateral and 60 recovery episodes — both targeting the v3 punch list surfaced by the v2 fine-tune's open-loop eval.

Renamed 2026-05-08: working name psi0-g1-sneaker-325ep-v1 → psi0-g1-sneaker-325ep-v3 to align the suffix with the fine-tuning round it feeds (Psi0 v3 fine-tune). The HF redirect from the old URL is preserved.

Quick facts

Episodes	325
Frames	111 443 @ 30 fps
Format	LeRobot v2.1
Camera	egocentric, 640×480, H.264 yuv420p
Robot	Unitree G1 + Inspire RH56DFTP (5-finger, 6 active DoF/side, padded to 14-D)
Action dim	36
State dim	32
Task instruction	"Pick up the sneaker and place it inside the box."
Episode length	min 5.9 s (177 f), mean 11.4 s (342 f), max 26.9 s (807 f)
Data collection	Pico 4 Ultra teleop with hand trackers (CloudWalk Robotics Lab)
License	Apache-2.0

Episode composition

episode_index range	Collection	Count	Description
0–93	v1 right-hand	94	RIGHT hand demos, standard placement (matches psi0-g1-sneaker-94ep-v1)
94–204	v2 left-hand	111	LEFT hand demos, standard placement (the LEFT half of 205ep-v2-source)
205–264	lateral	60	sneaker placed further to the sides — broadens RIGHT-side scene coverage v2 was thin on
265–324	recovery	60	demos including recovery from suboptimal grasps — first time this scene class appears in the dataset family

Sources (from meta/compose_manifest.json):

• sneaker_box_left_right (205 eps) → psi0-g1-sneaker-205ep-v2-source
• pick_up_the_sneaker_and_place_it_in_the_box_more_on_sides (60 eps)
• pick_up_the_sneaker_and_place_it_in_the_box_with_recovery (60 eps)

Why v3 (and what this dataset answers)

The v2 fine-tune (psi0-g1-sneaker-199ep-v2-finetune) showed two failure modes in its 2026-05-07 open-loop eval (full results on the v2 eval-set card):

1. Memorization — OOD/ID gap 5.5×–7.2× at ckpt_40000.
2. Per-side asymmetry — RIGHT-pickup hand_joints 3× worse than LEFT-pickup, despite an episode count split of only 1.19× LEFT-heavy. This pointed to RIGHT-side scene-level coverage (sneaker pose, lighting, distractors) as the bottleneck rather than raw episode count.

v3 is the dataset response: +60 lateral (broader sneaker pose coverage on both sides) +60 recovery (first-time exposure to suboptimal-grasp re-attempts, which v1 Phase 14 flagged as a real-robot failure mode and v2 never addressed). The third v3 priority — held-out OOD probe in the inner training loop — remains a code-side change, not a dataset change.

The 385ep v4 dataset further layers +60 negative episodes (no shoe / no box) on top of v3.

Action and state layout

Standard Psi0 36-D / 32-D layout (identical to the 199ep-v2 and 205ep-v2-source dataset families):

Range	Meaning
action[0:6]	LEFT hand 6 DoF
action[6]	LEFT hand pad — constant 0.0
action[7:13]	RIGHT hand 6 DoF
action[13]	RIGHT hand pad — constant 0.0
action[14:21]	LEFT arm 7 joints (rad)
action[21:28]	RIGHT arm 7 joints (rad)
action[28:31]	Torso RPY — constant 0
action[31]	Torso height — near-constant ≈ 0.749
action[32:36]	Locomotion (vx, vy, vyaw, target_yaw) — constant 0

Inspire RH56DFTP DDS command order is [pinky, ring, middle, index, thumb-bend, thumb-rotation] per side. Slots [6] and [13] are zero-padding to align with the canonical 14-D Psi0 hand layout. Predicted ill-conditioned action dims (verify post-download from meta/stats_psi0.json): {6, 13, 28, 29, 30, 32, 33, 34, 35} — same set as v2 (the lateral and recovery episodes preserve the bimanual + stationary base + Inspire-padding shape).

How to download

import os
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id='cloudwalk-research/psi0-g1-sneaker-325ep-v3',
    repo_type='dataset',
    local_dir='data/real/psi0-g1-sneaker-325ep-v3',
)

After download, run scripts/data/patch_lerobot_meta.py to strip stale huggingface metadata keys (required for datasets==3.6.0).

How to use for training

The v3 fine-tune launches Psi0 with:

bash scripts/train/psi0/finetune-real-psi0.sh \
    psi0-g1-sneaker-325ep-v3 sneaker-real-v3 \
    --train.max_training_steps=160000 \
    --train.checkpointing_steps=10000 \
    --train.max_checkpoints_to_keep=20

160 000 steps, batch 16/GPU × 8 GPUs (B200), bf16, cosine LR with 1e-4 peak, RTC enabled (max_delay=8), VLM frozen. Step count is 4× v2's 40 000 — deliberate, given v3's dataset is 1.6× larger and substantially more diverse. With 111 443 frames and effective batch 128, 160 000 steps cover ~184 effective epochs (vs v2's 84). The 16 retained checkpoints (ckpt_10000 through ckpt_160000, every 10k) are the diagnostic ladder that lets us locate the OOD optimum post-hoc without re-training.

Open-loop eval (v3 fine-tune)

Pending — to be done in a follow-up session, mirror of v2 Phase 11.5 (which uses scripts/eval/openloop_eval_psi0.py with --eval-dataset-id, commit f094c20). Recommended ckpts: 30k, 60k, 90k, 120k, 160k (5-point sweep across the 16 retained). Held-out eval split for v3 is itself a follow-up dataset-engineering task.

Lineage

psi0-g1-sneaker-94ep-v1            (v1 fine-tune dataset, RIGHT-hand only)
                ↓
psi0-g1-sneaker-205ep-v2-source    (canonical, immutable)
                ↓
        ├── psi0-g1-sneaker-199ep-v2          (v2 train; 6 ep held out)
        └── psi0-g1-sneaker-6ep-v2-eval       (v2 held-out OOD probe)
                ↓
psi0-g1-sneaker-325ep-v3   ← (this) +60 lateral, +60 recovery
                ↓
psi0-g1-sneaker-385ep-v4   (+60 negative — no shoe/box)

Related datasets

Repo	Episodes	Composition	Purpose
psi0-g1-sneaker-94ep-v1	94	RIGHT only	v1 fine-tune
psi0-g1-sneaker-205ep-v2-source	205	bimanual base	canonical source
psi0-g1-sneaker-199ep-v2	199	bimanual	v2 fine-tune training
psi0-g1-sneaker-6ep-v2-eval	6	3 L / 3 R	v2 held-out OOD eval
psi0-g1-sneaker-325ep-v3 (this)	325	bimanual + lateral + recovery	v3 fine-tune training
psi0-g1-sneaker-385ep-v4	385	+ 60 negative	v4 fine-tune training

Attribution

Developed by cloudwalk-research in the CloudWalk Robotics Lab (CW-RL). Part of the Psi0 collection. Built to fine-tune the Psi0 base model; stored in the LeRobot v2.1 format.

Citation

@misc{cwrl_psi0_g1_sneaker_325ep_v3_2026,
  title        = {Psi0 G1 Sneaker-in-Box --- 325 episodes (v3 training, bimanual + lateral + recovery)},
  author       = {{CloudWalk Robotics Lab}},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  url          = {https://huggingface.co/datasets/cloudwalk-research/psi0-g1-sneaker-325ep-v3}
}

@article{psi0_2026,
  title         = {Psi0: A Generalist Vision-Language-Action Model for Humanoid Robots},
  author        = {Wei, Songlin and others},
  year          = {2026},
  eprint        = {2603.12263},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2603.12263}
}

@article{black2024pi0,
  title   = {{$\pi_0$}: A Vision-Language-Action Flow Model for General Robot Control},
  author  = {Black, Kevin and Brown, Noah and Driess, Danny and others},
  journal = {arXiv preprint arXiv:2410.24164},
  year    = {2024}
}

@software{lerobot2024,
  title  = {{LeRobot}: State-of-the-art Machine Learning for Real-World Robotics in {PyTorch}},
  author = {{Hugging Face}},
  url    = {https://github.com/huggingface/lerobot},
  year   = {2024}
}