Downloads · 30 days
0
cloudwalk-research/psi0-g1-sneaker-94ep-v1-finetune
psi0-g1-sneaker-94ep-v1-finetune is a robotics model from cloudwalk-research. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for psi. The card lists the license as apache-2.0.
v1 community fine-tune of Psi0 (paper) on a Unitree G1 humanoid with Inspire RH56DFTP hands.
Downloads · 30 days
0
Access
Public
Updated May 15, 2026
Repo size
18.8 GB
Likes
0
Public
Click a slice to open those files.
.safetensors18.8 GB · 100%
From the Hugging Face model README
v1 community fine-tune of Psi0 (paper) on a Unitree G1 humanoid with Inspire RH56DFTP hands.
One-handed sneaker pick-and-place task (RIGHT-hand only). Trained at the CloudWalk Robotics Lab (CW-RL), 2026-04, as the first publicly reported external reproduction of the Psi0 paper's full pipeline (data collection → fine-tuning → real-robot deployment). Released as a reference fine-tune for teams using G1 + Inspire (instead of the upstream G1 + Dex3-1 baseline).
Successor: cloudwalk-research/psi0-g1-sneaker-199ep-v2-finetune — bimanual v2 trained on 199 episodes (108 LEFT + 91 RIGHT), addresses v1's single-handed limitation surfaced during real-robot deployment.
| Base | Psi0 (Qwen3-VL-2B + flow-based MM-DiT action expert) |
| Pretrained | pre.fast.1by1.2601091803.ckpt.ego200k.he30k + postpre.1by1.pad36.2601131206.ckpt.he30k |
| Dataset | cloudwalk-research/psi0-g1-sneaker-94ep-v1 (94 episodes, 26,192 frames @ 30 Hz, 640x480 ego cam) |
| Robot target | Unitree G1 + Inspire RH56DFTP (NOT Dex3-1) |
| Action space | 36-D (6 hand + 1 pad / side, 14 arm, 4 torso, 4 locomotion) |
| State space | 32-D (padded to 36-D for the model) |
| Action chunk | 30 |
| Hardware | 8x NVIDIA B200 (sm_100 / Blackwell) |
| Mixed precision | bf16 |
| Optimizer | AdamW, lr=1e-4 cosine + min_lr 5e-7, warmup 1000, betas=(0.95, 0.999), wd=1e-6, grad-clip 1.0 |
| Steps / batch | 40,000 / global batch 128 (16 per GPU x 8 GPUs) |
| RTC | enabled, max_delay = 8 |
| VLM | frozen (only action header trained) |
| Wall-clock | 3 h 58 min |
| Final eval/loss | 0.0675 (from 37.88 — ~560x) |
| W&B run | hjy0caz0 (CW-RL internal) |
Optimizer / scheduler / sampler / random-state files are intentionally omitted — needed only to resume training, would add ~3.7 GB per checkpoint.
Validation runs on a fresh-batch sample of the training repo (Psi0 standard practice — val_repo_ids defaults to train_repo_ids; the dataset is too small for a held-out split). Treat eval/loss as a smoothed train-loss probe, not as held-out generalization.
| Metric | Step 0 | Step 40k | Improvement |
|---|---|---|---|
| eval/loss | 37.88 | 0.0675 | ~560x |
| eval/err_l1_arm_joints | 3.40 | 0.058 | ~58x |
| eval/err_l1_hand_joints | 1.39 | 0.012 | ~115x |
| eval/err_l1_height | 5.0e-3 | 1.0e-5 | ~500x |
| eval/err_l1_{vx,vy,vyaw,target_yaw,torso_rpy} | 0 | 0 | task is stationary |
Full training curves tracked internally (W&B run hjy0caz0).
In-process evaluation against the training dataset, mirroring step 4 of the pre-deploy checklist Songlin posted on Psi0 issue #39: runs the trained model with the same image / state preprocessing pipeline used at inference and compares the predicted action chunk against the ground-truth action chunk. Catches transform / state-pad / image-preprocess regressions that the training-time eval/loss cannot, since training and validation share the same data pipeline by default.
Setup: 5 episodes (8, 40, 60, 70, 93) sampled deterministically with seed=42, stride 4, 10 diffusion inference steps, ~362 frames per checkpoint, single B200 GPU. Reference threshold from Songlin's reply (issue #39): hand and arm joints "usually around 0.12 rad" for a converged fine-tune.
| Group (denormalized L1 error norm) | ckpt_20000 | ckpt_30000 | ckpt_40000 | Songlin ref | Unit |
|---|---|---|---|---|---|
| hand_joints | 0.0309 | 0.0140 | 0.0117 | ~0.12 | rad |
| arm_joints | 0.1170 | 0.0738 | 0.0575 | ~0.12 | rad |
| rpy (torso) | 0.0000 | 0.0000 | 0.0000 | — | rad |
| height | 0.0000 | 0.0000 | 0.0000 | — | m |
| vx, vy, torso_vyaw, target_yaw | 0.0000 | 0.0000 | 0.0000 | — | m/s, rad |
Both hand_joints and arm_joints are well below Songlin's reference threshold at every retained checkpoint, with monotone improvement across ckpt_20000 → ckpt_30000 → ckpt_40000. The locomotion / torso dims at exactly 0 confirm the policy reproduces the stationary nature of the dataset (locomotion DoF are constant 0 in the GT actions; the model converges to that constant).
The script that produced this table is a parametrized port (openloop_eval_psi0.py, internal at CW-RL) of examples/simple/openloop_eval.ipynb from upstream Psi0.
import os
from huggingface_hub import hf_hub_download
repo = "cloudwalk-research/psi0-g1-sneaker-94ep-v1-finetune"
local_dir = ".runs/finetune/sneaker-in-box-real"
for f in ["run_config.json", "argv.txt", "checkpoints/ckpt_40000/model.safetensors"]:
hf_hub_download(repo_id=repo, filename=f, repo_type="model",
local_dir=local_dir)
bash sim/scripts/run_closed_loop.sh \
"Pick up the sneaker and place it inside the box." \
.runs/finetune/sneaker-in-box-real 40000 \
--fixed_base --live_viewer
cloudwalk-research/psi0-g1-sneaker-94ep-v1 (CloudWalk Research, 2026), Pico 4 Ultra teleoperation on G1 + Inspire RH56DFTP.cloudwalk-research/psi0-apple-to-plate-teleop (dataset).Developed by cloudwalk-research in the CloudWalk Robotics Lab (CW-RL). Part of the Psi0 collection. Fine-tuned from the Psi0 base model; integrates with the LeRobot framework.
If you build on this checkpoint, cite both upstream Psi0 and this fine-tune:
@misc{cwrl_psi0_sneaker_v1_2026,
title = {Psi0 Fine-Tune v1 --- G1 Sneaker-in-Box (Inspire RH56DFTP, right-hand only)},
author = {{CloudWalk Robotics Lab}},
year = {2026},
howpublished = {Hugging Face model repository},
url = {https://huggingface.co/cloudwalk-research/psi0-g1-sneaker-94ep-v1-finetune}
}
@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}
}
Released under Apache-2.0. This is a research preview: not intended for safety-critical use; closed-loop deployment on a physical humanoid requires human oversight.