PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking

Recent advances in vision--language--action and diffusion-based robot policies have largely come from training-side gains, but reliable deployment also depends on how predicted actions are executed. Under action chunking, each query predicts a sequence of future actions, and the robot executes an open-loop prefix before re-querying. The length of this prefix, the execution horizon, is a critical inference-time variable that trades off feedback frequency against motion continuity. Since fixed horizons exhibit strongly task-dependent and non-monotonic effects on success, no single constant horizon provides a reliable cross-task deployment rule. We propose PACE (Phase-Aware Chunk Execution), a training-free test-time execution method that selects the execution horizon online from the predicted chunk itself. PACE exploits the phase-dependent kinematic structure of manipulation trajectories, using prominent low-speed valleys in the predicted speed profile as candidate replanning points. On the 50-task RoboTwin2.0 benchmark, PACE improves the average success rate from 57.8\% to 64.2\% over the strongest fixed-horizon baseline, without task-specific fixed-horizon tuning. In real-robot experiments, PACE improves the average task score from 60.7 to 77.7 and the average success rate from 50.7\% to 70.4\%.