diantoudefengshan/Jetson-PI-pi05

Jetson-PI LIBERO Spatial Checkpoints

Pretrained checkpoints for paper Jetson-PI: Towards Onboard Real-Time Robot Control via Foresight-Aligned Asynchronous Inference on LIBERO benchmark, used with the Jetson-PI codebase.

This HuggingFace repo bundles two components (do not merge their params/ trees):

Subdirectory	Role	Size (approx.)
pi05_libero/	π₀.₅-LIBERO policy backbone (VLM + action expert)	~12 GB
future_correction_module/	Trained future correction module (WM step 65000)	~157 MB

Training summary

• Policy: Physical Intelligence pi05_libero (included for convenience; see license below).
• Future correction module: three-stage FAAC training on LIBERO-spatial (task_index 30–39), global step 65000.
  • Stage 1: 30k (L_act, action expert + token reducer)
  • Stage 2: 15k (L_cond, future correction module)
  • Stage 3: 55k (joint L_cond + L_act, μ detached)
• Inference defaults: H=10, K=9, action_encoder=transformer_block, token_reducer=learned_cross_attn.

Directory layout

.
├── README.md
├── config.json
├── pi05_libero/
│   ├── assets/physical-intelligence/libero/norm_stats.json
│   └── params/                 # Orbax π₀.₅ weights
└── future_correction_module/
    └── params/                 # Orbax future correction module weights

Usage with Jetson-PI

After cloning the code repo and installing dependencies:

export PI0_CHECKPOINT=/path/to/hf_snapshot/pi05_libero
export WM=/path/to/hf_snapshot/future_correction_module
export PY_SERVER=/path/to/python_with_jax
export PY=/path/to/python

# Evaluation (LIBERO-spatial)
bash scripts/eval_wm_libero_spatial.sh

Or point serve_policy.py directly:

python scripts/serve_policy.py --env LIBERO --port 8000 \
  --world-model-checkpoint "${WM}" \
  --world-model-token-reducer-kind learned_cross_attn \
  --world-model-action-encoder-kind transformer_block \
  --async-ae-proprio-source prefix_t \
  policy:checkpoint --policy.config pi05_libero --policy.dir "${PI0_CHECKPOINT}"

Download

pip install huggingface_hub
huggingface-cli download <your-org>/jetson-pi-libero-spatial --local-dir ./jetson-pi-libero-spatial

Citation

If Jetson-PI or Jetson-PI-Edge helps your research, please cite our paper:

@article{yang2026jetson,
  title={Jetson-PI: Towards Onboard Real-Time Robot Control via Foresight-Aligned Asynchronous Inference},
  author={Yang, Zebin and Wang, Qi and Wang, Yunhe and Guo, Xiurui and Yu, Bo and Liu, Shaoshan and Xu, Jiafeng and Dong, Hao and Li, Meng},
  journal={arXiv preprint arXiv:2607.12659},
  year={2026}
}

License

• pi05_libero/ weights are from Physical Intelligence openpi (π₀.₅ + Gemma). See LICENSE_GEMMA.txt in the code repo and the openpi license.
• future_correction_module/ is released under the same terms as the Jetson-PI project.