pi0-fast-primitive-deltachunk-w-two-bridge

Pi0-fast official implementation trained on VLABench datasets.

This repository provides the official release of the Pi0-fast model trained with the whole VLABench official primitive tasks dataset. This config corresponds to the delta chunk setting and uses two bridge co-training data streams.

The uploaded checkpoint includes both inference parameters and the full training state:

• params/: Orbax parameters for inference/evaluation.
• train_state/: full Orbax training state for resuming training.
• assets/: normalization statistics and checkpoint assets.
• _CHECKPOINT_METADATA: Orbax checkpoint metadata.

Evaluation

To run this checkpoint, please clone this repo: https://github.com/Shiduo-zh/openpi, and checkout to the branch main. Assume that you download this checkpoint and put it in the directory checkpoints, to run the policy as server, please run:

bash vla_bench_scipts/serve_policy.sh pifast_w_vlabench_delta_cotrain_eb checkpoints/VLABench/pi0-fast-primitive-deltachunk-w-two-bridge/99999/

After serving the policy, open another terminal and run:

bash vla_bench_scipts/multi_run_vlabench.sh <Your path to store the evaluate results>

Train

To reproduce the training result, please run the training script with the config pifast_w_vlabench_delta_cotrain_eb.

XLA_PYTHON_CLIENT_MEM_FRACTION=0.95 uv run scripts/train_cotrain.py pifast_w_vlabench_delta_cotrain_eb \
  --exp-name=pifast_w_vlabench_delta_cotrain_eb \
  --batch-size=32 \
  --save_interval=10000 \
  --overwrite

Our checkpoint is trained for 100k iterations on the VLABench primitive tasks dataset.

Reference Results

The reference success rate of this model is:

Track	add_condiment	insert_flower	select_book	select_chemistry_tube	select_drink	select_fruit	select_mahjong	select_painting	select_poker	select_toy	Avg_SR
track_1_in_distribution	0.32	0.26	0.837	0.86	0.42	0.84	0.72	0.56	0.94	0.72	0.648
track_2_cross_category	0.08	?	0.122	0.18	0.08	0.7	0.51	0.34	0.96	0.62	0.399

Citation

If you use this checkpoint, please consider to cite:

@article{yin2026two,
  title={Two Bridges, One Pathway: From VLMs to Generalizable VLAs with Embodied Trajectory-Coupled Data},
  author={Yin, Linqi and Zhang, Shiduo and Qiu, Shenling and Li, Chenxin and Fu, Zhaoyang and Xiao, Lei and Wang, Xiang and Yang, Chenchen and Xu, Zhe and Qian, Pengfang and others},
  journal={arXiv preprint arXiv:2606.08520},
  year={2026}
}

@article{zhang2024vlabench,
  title={Vlabench: A large-scale benchmark for language-conditioned robotics manipulation with long-horizon reasoning tasks},
  author={Zhang, Shiduo and Xu, Zhe and Liu, Peiju and Yu, Xiaopeng and Li, Yuan and Gao, Qinghui and Fei, Zhaoye and Yin, Zhangyue and Wu, Zuxuan and Jiang, Yu-Gang and others},
  journal={arXiv preprint arXiv:2412.18194},
  year={2024}
}