Development of a Novel Compact Robotic Exoskeleton Glove with Reinforcement Learning Control

Abstract

This paper presents the design, optimization, control, and experimental evaluation of a novel compact exoskeleton glove aiming to assist patients with brachial plexus injuries in grasping daily used objects. The finger mechanism is based on a rigid coupling hybrid mechanism (RCHM) concept, which utilizes a serially connected rack-and-pinion mechanism and an offset slider-crank mechanism to couple the motions of different finger joints. The glove dimensions are synthesized based on the natural grasping motion of human hands. To better control the glove and enhance the grasping capabilities, a simulation environment was developed and reinforcement learning techniques were applied. This learning-based control trained an agent to perform different grasp types with appropriate force. The trained agent was then applied in real-world experiments with the developed exoskeleton glove. The results validated the effectiveness of the mechanical design and the real-time self-adjustable control policy, which demonstrated the glove's functionality and capability to grasp various objects relevant to activities of daily living (ADLs)