Trajectory Planning of Rehabilitation Exercises using an Integrated Reward Function in Reinforcement Learning

Abstract

Introduction: Rehabilitation devices help patients to recover injured body parts such as elbow and knee joints [3]. Trajectory planning of rehabilitation exercises determines a suitable moving path to guide patients in daily recovery activities for body parts based on injured levels and joints [4]. It is expected that the rehabilitation process is smooth and comfortable. The existing trajectory planning are mainly manual methods that require physicians to plan the rehabilitation exercise trajectory [7], which is inefficient and inaccurate [1]. Reinforcement learning (RL) uses intelligent agents to plan actions in environments for maximum rewards [5]. Using RL, a rehabilitation device can autonomously learn and plan a trajectory for required exercise actions in different conditions. Based on the range of rotation angles and movement speed required in the rehabilitation of patients, a reward function can generate the optimal trajectory for patients to approach the target position in rehabilitation exercises efficiently and accurately [6]. An integrated reward function is proposed in this paper to plan the trajectory of rehabilitation exercises. Based on injured joints of a patient recorded by motion sensors, the range of rotation angles and movement speeds are restricted and planed for the patient using RL. The rotation angles and movement speeds are reset for injured joints based on the daily progress of the patient recovery to improve performance of the rehabilitation.