DexPolicy: Scheduled Exploration for Trajectory-Guided Dexterous Manipulation
DexPolicy anneals exploration for dexterous manipulation, raising target success for PPO, GRPO, and FPO.
Dexterous-manipulation reinforcement learning must discover finger-object contacts and then control the object, but persistent action noise can interfere with precise control. DexPolicy makes exploration scale a function of training steps, annealing from broad to narrow while holding loss, architecture, reward, and optimizer fixed. Across five YCB objects, mean deterministic Target success rises from 49.4% to 68.1% for flow-parameterized PPO, 14.1% to 45.4% for GRPO, and 32.0% to 35.7% for PPO. On a RealMan RM75 arm with an Inspire RH56 hand, 360 trials raise FPO success from 25.0% to 85.0%.
- Exploration scale anneals from broad to narrow with other training settings fixed.
- Simulation FPO Target success rises from 49.4 percent to 68.1 percent.
- Real-robot FPO success rises from 25.0 percent to 85.0 percent over 360 trials.
- Terminal task success, not training return, should select exploration schedules.
- Products
- RealMan RM75Inspire RH56
From the original article
1 min read at huggingface.co
Reinforcement learning (RL) for dexterous manipulation must discover finger-object contacts and then control the object precisely; the action noise that serves the first goal can interfere with the second.
By Haoyu Wang, Siyuan Qian, Yanjun Li · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.