Real-Time Execution with Autoregressive Policies

This research paper presents a method for achieving real-time execution in large-scale Vision-Language-Action models using autoregressive policies. By adjusting the tokenization horizon and applying constrained decoding, the authors ensure smooth action trajectories and fast reactivity, overcoming the slower rollout speed typically associated with synchronous inference in autoregressive models. The study shows that these policies consistently outperform equivalent flow-matching policy counterparts across both simulated and real-world environments, achieving significantly improved task completion speeds. Additionally, the inherent advantages of autoregressive policies, such as faster convergence and better generalizability in instruction-following, are highlighted, confirming their competitiveness for supporting real-time execution.