FlashVLA · π0.5 · LIBERO

Streaming Action Decoding for Fast and Asynchronous VLA Inference

A π0.5 flow-matching vision-language-action policy finetuned on LIBERO 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: LIBERO (Spatial / Object / Goal / Long)

Results

LIBERO success rate (%) and per-step latency. d is the async step delay: d=0 is synchronous, d>=1 overlaps the next chunk's inference with robot execution.

Model	Spatial	Object	Goal	Long	Avg	Time/Step (ms)
π0.5 (base)	98.8	98.2	98.0	92.4	96.9	53.8
+FlashVLA (`d=0`)	98.6	99.0	97.8	96.2	97.9	—
+FlashVLA (`d=1`)	96.0	99.6	96.4	96.0	97.0	29.4 (1.83× faster)

Usage

Install FlashVLA:

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

LIBERO evaluation needs additional simulator-specific setup, described in sim_eval/libero/. Once that is in place:

bash sim_eval/libero/eval.sh

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

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}
}