Asynchronous control speeds up robots driven by vision-based policies

Takayuki Osa and Kazuki Ota have presented a framework that lets robots carry out tasks faster with vision-based control policies, including vision-language-action models, without changing how the policies are trained, in a paper for the Robot Learning Workshop at NeurIPS 2026.

Such policies are often slowed by the computing cost of choosing each action, the authors write. Their framework runs inference and control asynchronously and treats turning the model's separate predictions into continuous target positions as a tracking control problem, which keeps the robot's motion smooth and stable. It speeds execution further by skipping actions when the variance of predicted positions allows.

Experiments with the Action Chunking Transformer and OpenVLA-OFT policies on a robot with two arms showed substantial speedups while task success stayed high, the authors report.