A Learning Framework for Dexterous Manipulation Skills in Multi-Finger Robot Hands

Due to the high-dimensionality of the control space and environmental unpredictability, dexterous manipulation by robots has always been a challenging issue. It is essential for robots to possess dexterous manipulation skills, such as engaging in household chores in a home environment. To address this challenge, we propose a learning framework aimed at advancing complex manipulation skills in multi-finger robot hands. Specifically, we leverage keypoints to simplify the success criteria and introduce a fingertip contact reward to promote precise fingertip-object interactions. The generalizability of our policy is assessed by using unseen objects of the same category in the same scenario, and its robustness is further tested by integrating domain randomization training. Our framework is tested on the Isaac Gym platform with a 16-degree-of-freedom Allegro Hand and a 7-degree-of-freedom KUKA IIWA robotic arm.