Reflex: Fast Robot Policy Serving by Reusing and Correcting Past Actions
This research presents Reflex, a new serving layer for robot policies that addresses latency and throughput bottlenecks by caching complete action chunks rather than intermediate computations. The method employs a two-level correction mechanism to handle observation-action mismatches and accumulated errors without modifying the underlying policy. Evaluated on representative workloads and real robots, Reflex improves inference latency by 1.4x to 11x and throughput by 2.0x to 2.2x while preserving comparable task success rates.