SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics

1 min read

SmolVLA is a small vision-language-action model meant to be trained on one GPU and run on consumer GPUs or even CPUs. It is trained on community-collected data from affordable robot platforms, and pairs chunked action generation with an asynchronous inference stack that separates perception and action prediction from action execution, allowing higher control rates.

Findings

  • Performance comparable to VLAs about 10 times larger, on simulated and real-world benchmarks (the abstract gives no success rates).
  • Code, pretrained models and training data are released.

Limits

  • Abstract-level reading only.

Relevance

A VLA small enough to run on board suits robots with weak links to a base station, where offloading to a large model is not an option. Asynchronous inference also means the robot acts on observations that are slightly old; how stale they can get before that becomes a problem is an open question for safe-control and robust robotics. Part of LeRobot.