DRC: pre-actuated dynamic residual control for high-speed rigid interception

Résumé fourni par la source

Mobile manipulation is increasingly transitioning from quasi-static settings to highly dynamic interaction scenarios, where millisecond-level responsiveness becomes critical. However, most existing control paradigms assume ideal instantaneous actuation and primarily emphasize perception accuracy, overlooking the impact of actuator dynamics and communication latency. This mismatch introduces a pronounced execution gap in real-world systems, resulting in severe phase lag during high-frequency tasks. To address this challenge, Dynamic Residual Control (DRC) is presented as a hybrid control framework that decouples global kinematic guidance from execution-level dynamic compensation. A frozen Nominal Kinematic Policy provides ideal kinematic commands, while a lightweight reinforcement learning–based residual policy operates in parallel at the control interface to correct high-frequency execution errors. Robust Sim-to-Real transfer is achieved by explicitly incorporating action history to capture inertial effects and by applying aggressive Dynamics Domain Randomization (DDR), which induces a pre-actuation behavior that counteracts system latency. Real-world experiments on a rigid non-prehensile interception benchmark demonstrate that DRC reduces end-to-end response latency from 120 ms to 35 ms and improves interception success rates from 72.0% to 83.0%, without relying on mechanical compliance or hardware modification.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.