kehong/FastVLA-weights

Fast-VLA (FastVLA) checkpoints

Paper (arXiv:2610.21389) · Project page · Code

Fast-VLA calls a frozen 4B vision-language model (DM0.5) once per episode and keeps its layer-wise key/value cache as a fixed task memory. At every action chunk only a lightweight online actor runs: it reads the current views and robot state, queries the task memory, and predicts the next chunk of actions.

This repository holds the online actors (DINOv2-based encoders + action decoder). The frozen DM0.5 VLM is downloaded from the Hugging Face Hub on first use.

Folder	Benchmark	Trained on	Parameters	Success rate (%)
fastvla-robotwin/	RoboTwin 2.0	27,500 demos (50 clean + 500 randomized per task)	272M	92.46 Clean / 92.40 Randomized
fastvla-libero/	LIBERO	2,000 demos (Spatial / Object / Goal / Long)	185M	95.95 LIBERO / 79.11 LIBERO-Plus

Each folder contains model.safetensors, config.json, norm_stats.json and rope_tables.pt.

Usage

git clone https://github.com/Garfield-kh/FastVLA && cd FastVLA
hf download kehong/FastVLA-weights --local-dir checkpoints   # or --include "fastvla-libero/*"

from fastvla.policy import FastVLAPolicy

policy = FastVLAPolicy("checkpoints/fastvla-robotwin")
policy.reset()                                           # start of an episode
actions = policy.act([head_rgb, left_wrist_rgb, right_wrist_rgb], joint_targets,
                     "Place the cans in the plastic box.")   # (50, 14)

Installation, evaluation and training instructions are in the GitHub repository.

Citation

@article{gong2026fastvla,
  title   = {Fast-VLA: You Only Think Once for Robot Manipulation},
  author  = {Gong, Kehong and Hou, Guanli and Zheng, Xiaoxu and Wang, Qi and Xu, Mingxi and
             Zhai, Anbang and He, Xiaoyu and Zhang, Mingyuan and Mi, Michael Bi},
  journal = {arXiv preprint arXiv:2610.21389},
  year    = {2026}
}

License

The checkpoints are released under Apache-2.0. DM0.5, which the policy uses frozen, is distributed separately under its own license.