What changed
Researchers released FastVLA code and weights in October 2026. The work is on arXiv 2610.21389. The GitHub repo is Garfield-kh/FastVLA.
The policy calls a frozen 4B vision language model once per episode. That first pass reads the language goal and the opening camera views. It stores layer key and value caches as fixed task memory. After that, a smaller online actor runs each control step. It reads fresh views and robot state, queries the cache, and outputs the next action chunk.
The README reports strong sim scores on RoboTwin 2.0 and LIBERO while cutting per step inference time versus several prior VLA baselines on the same hardware. Checkpoints fastvla-robotwin and fastvla-libero sit on Hugging Face under kehong/FastVLA-weights. A project page hosts fuller benchmark tables and figures.
The repo ships a policy server plus clients for named simulators. Training code includes a first frame cache builder the authors call a bank. Ascend NPU install notes appear in the install doc for teams outside NVIDIA CUDA.
Why it matters
Robot policies that rerun a full VLM every step burn GPU time and memory. That cost blocks tight control loops on real arms. One heavy think per episode keeps language grounding while the online head stays light. Teams planning edge deploys can trade an upfront cache build for steady low latency steps.
The design also separates what must stay frozen from what you can fine tune on plant data. That split can simplify safety review when the base VLM weights stay vendor locked.
Who is affected
Robotics ML engineers building manipulation stacks on VLAs should compare this pattern to their current per step encoders. Product leads who set latency budgets for pick and place cells care about the reported millisecond per step figures. Sim benchmark users on RoboTwin 2.0 or LIBERO can reproduce numbers with the shipped policy server.
Hardware teams sizing GPUs for dual sim plus serve hosts should read resident memory numbers in the README table. Teams on Ascend should read the NPU branch before they promise CUDA only timelines.
What to do next
Clone the repo and follow the install block for Python 3.11 and the pinned torch build. Download one checkpoint and run the policy server against the sim client your lab already uses. Log end to end latency on your GPU, not only the README table.
If you fine tune, note that the VLM stays frozen. Plan data collection around the online actor and cache builder rather than full model updates. Capture the instruction text and first frame set you used to build the bank so runs stay repeatable.
What to watch
Watch for real robot demos beyond the sim clients named in the README. Watch whether other groups publish the same once per episode cache design on new benchmarks. Watch Hugging Face for updated weights or license notes on the frozen DM0.5 backbone.
Watch power draw on edge boxes when the cache lives in memory for long episodes. That metric often decides factory pilots more than sim success rate alone.
Teams comparing FastVLA to per step VLAs should align action chunk sizes and camera counts before they trust speedup ratios from the README. The repo documents socket based sim clients, so network hop latency on your laptop can skew small bench runs. Log both policy server GPU use and sim step time in one chart for fair plots.
Safety reviewers should ask what happens when the first frame cache is built from a mislabeled instruction. Because the heavy VLM runs once, a bad prompt may persist for the whole episode unless the stack rebuilds the bank on goal change.
Benchmark authors should cite the exact checkpoint name when they compare to FastVLA numbers. Robotwin and libero weights differ in size and training data mix per the Hugging Face tree.
The October news line in the README marks the first public code drop alongside arXiv. Teams waiting for real hardware evals should track the project page for any addendum figures.
Photo: ThisIsEngineering / Pexels
Sources
- Primary — Garfield-kh et al., FastVLA repository README (October 2026). One VLM call per episode, benchmarks, install, and checkpoint names.
- Primary — Kehong et al., arXiv 2610.21389 (October 2026). Paper listing for Fast VLA.
- Primary — Kehong, FastVLA weights on Hugging Face (October 2026). Model artifacts for robotwin and libero runs.
Related in Robotics
Continue reading