DexPolicy: Scheduled Exploration for Trajectory-Guided Dexterous Manipulation

DexPolicy addresses exploration-exploitation trade-offs in trajectory-guided dexterous manipulation by annealing action noise as a function of training steps. This method prevents broad exploration noise from destabilizing precision control, outperforming baseline PPO across five YCB objects using PPO, critic-free GRPO, and flow-parameterized PPO variants.