efficient inference for vision-language-action robot policiesIntermediate
Fast-VLA: You Only Think Once for Robot Manipulation
Huawei researchers run their 4-billion-parameter vision-language model only once per task, at the start, and let a small actor network consult that frozen memory for every later move. In simulation benchmarks this keeps success near the best models while cutting average inference time per control step by 10.9 to 17.1 times and needing about 1.2 GB of memory during execution.