Within-Hand Manipulation with an Underactuated Dexterous Hand Based on Pre-Trained Reinforcement Learning

This paper investigates a dexterous manipulation method for multi-fingered hands based on the approximate policy optimization (PPO) algorithm. First, a prior knowledge base of dexterous manipulation demonstration data is provided for the PPO algorithm, and the original model is pre-trained to guide fast network training. Second, a new reward function is proposed to optimize rotational manipulation action decisions based on hand joints, fingertip tactile, and object pose infor mation. Finally, a three-fingered hand is used as an example for rotational manipulation experiments, and the comparative simulation results show that the method is more efficient than the traditional PPO algorithm and provides ideas for dexterous manipulation of multi-fingered hands.