Dynamic Execution Horizon Prediction: Freezing the Policy While Learning When to Replan
The paper proposes DEHP, a method that trains a lightweight execution-horizon prediction branch via online reinforcement learning, keeping the pretrained chunk-based policy completely frozen. This allows dynamic adjustment of how many action steps the robot executes open-loop before replanning, rather than relying on a fixed task-dependent horizon. DEHP significantly improves success rates on high-precision and long-horizon tasks by predicting shorter horizons during fine-grained stages and longer horizons during free-space motion.