Datasets:

cloudwalk-research
/

psi0-g1-sneaker-205ep-v2-source

⚠️ Do not use this dataset directly for training. This is the canonical immutable union of the v1 and v2 collections, kept as a source of truth for reproducibility. For v2 fine-tuning use psi0-g1-sneaker-199ep-v2; for held-out evaluation use psi0-g1-sneaker-6ep-v2-eval. Together these two derivatives reconstruct this canonical dataset exactly: 199 + 6 = 205.

205 teleoperated episodes of a Unitree G1 humanoid (with Inspire RH56DFTP hands) picking up a sneaker and placing it inside a box. Bimanual coverage: roughly half the episodes use the left hand, half the right. Built from the v1 collection (94 right-hand episodes, indices 0–93) plus a new 111-episode left-hand pass (indices 94–204).

Quick facts


Episodes	205
Frames	62 447 @ 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)
Hand-side composition	111 LEFT (idx 94–204) + 94 RIGHT (idx 0–93)
Action dim	36
State dim	32
Task instruction	`"Pick up the sneaker and place it inside the box."`
Episode length	min 178, median ~ 275, max ~ 470 frames
Data collection	Pico 4 Ultra teleop with hand trackers (CloudWalk Robotics Lab)
License	Apache-2.0

Hand-side composition

`episode_index` range	Side	Count	Source
0–93	RIGHT	94	identical to `psi0-g1-sneaker-94ep-v1` (the v1 collection)
94–204	LEFT	111	new bimanual extension collected for v2

Per-episode classification (no per-episode label is stored in metadata; the side can be recovered from the action tensor):

left_unused = mean(action[:, 0:6]), right_unused = mean(action[:, 7:13])
side = "LEFT" if left_unused < right_unused else "RIGHT"
All 205 episodes are unambiguous (minimum margin 0.27 in [0, 1] hand-command space).

Action and state layout

Standard Psi0 36-D / 32-D layout. Dataset-specific behaviour for the canonical v2 source:

Range	Meaning	This dataset
`action[0:6]`	LEFT hand 6 DoF	active in 111 episodes (LEFT-side demos), constant `1.0` in 94 (RIGHT-side); aggregate `std ≈ 0.42`
`action[6]`	LEFT hand pad	constant `0.0`
`action[7:13]`	RIGHT hand 6 DoF	active in 94 episodes (RIGHT-side demos), constant `1.0` in 111 (LEFT-side); aggregate `std ≈ 0.37`
`action[13]`	RIGHT hand pad	constant `0.0`
`action[14:21]`	LEFT arm 7 joints (rad)	active in all episodes
`action[21:28]`	RIGHT arm 7 joints (rad)	active in all episodes
`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` (stationary task)

Within any single episode the hand commands are still trigger-broadcast (Pico 4 Ultra trigger drives all 6 active DoF in lockstep — single grip-strength scalar). The bimanual signal lives across episodes, not within them.

The Inspire RH56DFTP DDS command order (used by the deploy-side bridge) is [pinky, ring, middle, index, thumb-bend, thumb-rotation] per side. The 7th index per hand ([6] and [13] in the 14-D layout) is constant zero padding that aligns the dataset with the canonical 14-D Psi0 hand layout.

How to use

This canonical dataset is not plugged into the v2 training pipeline. The two derivatives are:

Repo	Episodes	Frames	Hand split	Purpose
`psi0-g1-sneaker-199ep-v2`	199	60 656	108 L / 91 R	training input for v2 fine-tune
`psi0-g1-sneaker-6ep-v2-eval`	6	1 791	3 L / 3 R	held-out open-loop eval (orig idx `[41, 56, 91, 101, 162, 181]`)

Each derivative renumbers episode_index to 0..N-1 and recomputes meta/stats.json / meta/stats_psi0.json over its own frames, so they are fully self-contained.

How to download

import os
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id='cloudwalk-research/psi0-g1-sneaker-205ep-v2-source',
    repo_type='dataset',
    local_dir='data/real/psi0-g1-sneaker-205ep-v2-source',
)

Run scripts/data/patch_lerobot_meta.py after download (strips stale huggingface metadata; required for datasets==3.6.0).

Provenance

94 right-hand episodes (idx 0–93): collected for v1 fine-tune (CloudWalk Robotics Lab, 2026-04). Identical bytes to psi0-g1-sneaker-94ep-v1.
111 left-hand episodes (idx 94–204): collected 2026-05 to test the bimanual hand-selection hypothesis — given roughly balanced LEFT/RIGHT data, can the policy learn to visually identify which side the shoe is on and pick the correct hand?

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_205ep_v2_source_2026,
  title        = {Psi0 G1 Sneaker-in-Box --- 205 episodes (v2 canonical source)},
  author       = {{CloudWalk Robotics Lab}},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  url          = {https://huggingface.co/datasets/cloudwalk-research/psi0-g1-sneaker-205ep-v2-source}
}

@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}
}