cloudwalk-research/psi0-g1-sneaker-385ep-v4

Psi0 G1 Sneaker-in-Box — 385 episodes (v4 training)

385 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 psi0-g1-sneaker-325ep-v3 with 60 negative episodes — scenes where the box and/or sneaker are absent — to teach the policy to remain still when the task cannot be executed.

Renamed 2026-05-08: working name psi0-g1-sneaker-385ep-v1 → psi0-g1-sneaker-385ep-v4 to align the suffix with the fine-tuning round it feeds (Psi0 v4 fine-tune). The HF redirect from the old URL is preserved.

Quick facts

Episodes	385
Frames	119 043 @ 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 3.0 s (180 f), mean 10.3 s (309 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
0–59	negative	60	scene without box and/or sneaker — policy should remain still
60–153	v1 right-hand	94	RIGHT hand demos, standard placement
154–264	v2 left-hand	111	LEFT hand demos, standard placement
265–324	lateral	60	sneaker placed further to the sides
325–384	recovery	60	demos including recovery from suboptimal grasps

Sources (from meta/compose_manifest.json):

• pick_up_the_sneaker_and_place_it_in_the_box_no_box_andor_sneakers (60 eps) — negative examples
• sneaker_box_325ep_v1 (325 eps) → psi0-g1-sneaker-325ep-v3 (post-rename)

Negative episodes

The 60 negative episodes (idx 0–59) use the same task instruction as positives ("Pick up the sneaker and place it inside the box."). During collection the operator kept the robot stationary while the scene was missing one or both objects. The intended effect is that the policy learns to condition on visual presence of the objects rather than execute the task blindly upon receiving the instruction — directly addressing v1 Phase 14's "near-identical motion when the shoe was removed" failure mode that v2 never closed.

Episode lengths in this subset are short on average (mean ~3 s) since there is no manipulation to perform. Frame-weighted exposure of negatives in the dataset is therefore lower than episode-weighted: 60/385 ≈ 16% of episodes but only ~5% of frames. This frame-vs-episode imbalance is a hyperparameter to watch in v4 fine-tune evaluation; if the policy still moves blindly with shoe absent, it may be a sign that the negative scenes are under-represented at the chunk level.

Why v4 (and what this dataset answers)

v4 closes the third item from v1 Phase 14's punch list (negative scenes) — the carryover that v2 never addressed and v3 only partially handled (recovery, but not no-object scenes). Together with v3's lateral + recovery additions, v4 gives the policy:

• Hand-side selection (from v2's 50/50 LEFT/RIGHT)
• Lateral scene coverage (from v3's more_on_sides)
• Failure-recovery exemplars (from v3's recovery)
• No-go cue (from v4's no_box_andor_sneakers) — the policy should learn that absence of the visual targets means stay-still, even when the language instruction says otherwise

Falsifier (real-robot relevant): if the v4 policy still moves identically with shoe absent, the negatives didn't bind. Likely cause: frame-count imbalance (~5% by frame, see above). Mitigation, if observed, is either to lengthen the negative episodes during collection or to up-sample them at training time.

Action and state layout

Standard Psi0 36-D / 32-D layout (identical to the 325ep-v3 / 199ep-v2 / 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

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 / v3. The 60 negatives lower aggregate hand-std modestly but do not drive any active dim to zero (the 325 positive episodes still cover both LEFT and RIGHT hands).

How to download

import os
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id='cloudwalk-research/psi0-g1-sneaker-385ep-v4',
    repo_type='dataset',
    local_dir='data/real/psi0-g1-sneaker-385ep-v4',
)

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 v4 fine-tune launches Psi0 with:

bash scripts/train/psi0/finetune-real-psi0.sh \
    psi0-g1-sneaker-385ep-v4 sneaker-real-v4 \
    --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. With 119 043 frames and effective batch 128, 160 000 steps cover ~172 effective epochs. The 16 retained checkpoints (ckpt_10000 through ckpt_160000, every 10k) are the diagnostic ladder for post-hoc OOD-optimum location.

Open-loop eval (v4 fine-tune)

Pending — to be done in a follow-up session. Recommended ckpts: 30k, 60k, 90k, 120k, 160k. Same eval pipeline as v2 / v3 (scripts/eval/openloop_eval_psi0.py with --eval-dataset-id).

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            (+60 lateral, +60 recovery)
                ↓
psi0-g1-sneaker-385ep-v4   ← (this) +60 negative (no shoe / no 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	325	+ lateral + recovery	v3 fine-tune training
psi0-g1-sneaker-385ep-v4 (this)	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_385ep_v4_2026,
  title        = {Psi0 G1 Sneaker-in-Box --- 385 episodes (v4 training, bimanual + lateral + recovery + negative)},
  author       = {{CloudWalk Robotics Lab}},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  url          = {https://huggingface.co/datasets/cloudwalk-research/psi0-g1-sneaker-385ep-v4}
}

@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}
}