FlashVLA · π0.5 · RoboTwin 2.0

Streaming Action Decoding for Fast and Asynchronous VLA Inference

A π0.5 flow-matching vision-language-action policy finetuned on RoboTwin 2.0 (50-task multitask) and served with FlashVLA streaming action decoding for fast, asynchronous inference.

• Base model: lerobot/pi05_base
• Method: FlashVLA, streaming action decoding for flow-matching VLAs (async chunk-overlap execution)
• Benchmark: RoboTwin 2.0, 50-task multitask (clean / randomized)

Results

RoboTwin 2.0 50-task multitask success rate (%). d is the async step delay: d=0 is synchronous, d=1 and d=2 overlap the next chunk's inference with execution.

Usage

Install FlashVLA:

git clone https://github.com/z-lab/flashvla.git
cd flashvla
conda env create -f environment.yml
conda activate flashvla

RoboTwin 2.0 evaluation uses a server/client split across two environments. After the one-time setup in sim_eval/robotwin/, from $ROBOTWIN/policy/pi05_flashvla/:

bash eval_server.sh                      # terminal 1, flashvla env, starts the policy server
ROBOTWIN_VENV=... bash eval_client.sh    # terminal 2, RoboTwin env, runs the SAPIEN sim

Training configs for this checkpoint are in train/configs/pi05/robotwin/.

License

These weights are finetuned from lerobot/pi05_base, which is released under the Gemma Terms of Use. Those terms govern model derivatives, so they apply to this checkpoint and to anything derived from it, including the Gemma Prohibited Use Policy. If you redistribute this checkpoint or a derivative of it, you must pass the same terms along.

The FlashVLA inference and training code is separately released under the Apache 2.0 License.

Citation

@inproceedings{li2026flashvla,
  title     = {{FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference}},
  author    = {Li, Zekai and Tang, Jiaming and Liu, Zhijian},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026}
}