REFINE-DP presents RL fine-tuning method for data-efficient humanoid loco-manipulation

Evidence notes

• REFINE-DP jointly fine-tunes a Diffusion Policy planner and an RL-based whole-body controller to reduce distribution mismatch between high-level planning and low-level execution.
• The method reports over 90 percent task success in simulation and validates robust real-world execution on a 29-DoF Booster T1 humanoid across box pickup, object transport, door opening, and long-horizon manipulation tasks.
• The paper reports that a policy pre-trained on about 50 teleoperated trajectories can reach up to 95 percent success after fine-tuning, compared with roughly 1,000 trajectories needed to reach 90 percent success through pre-training alone.