Towards End-to-End Control of a Robot Prosthetic Hand via Reinforcement Learning

Abstract

Robot prosthetic hands intend to replicate one's lost abilities through intuitive control. So far, control methods that rely heavily on the human input such as Electromyographic (EMG) and Electroneurographic (ENG) signals have been predominantly studied. However, these methods face issues such as lack of robustness resulting in abandonment of this technology by the users. There is a need for a paradigm shift in the robot prosthetic hand control methods. With this regard, we propose an end-to-end learning of control policy for a robot prosthetic hand through reinforcement learning. Imitation learning has been fostered to help with the sparse reward setting in the hard-to-explore state-space of the problem. The results in simulation show the feasibility of successfully learning an endto-end policy for grasping objects by robot prosthetic hands, potentially increasing robustness for grasp control of future robot prosthetic hands.