Deep Reinforcement Learning Based Cable Tension Distribution Optimization for Cable-driven Rehabilitation Robot

Cable-driven parallel robots (CDPRs) have become increasingly widely used in rehabilitation training, while the control of cable tension remains a major problem. In this study, we propose a method based on deep reinforcement learning to optimize cable tension distribution in a 3-DOF CDPR, which has one redundant cable and is used for upper-limb rehabilitation. The reference trajectory with an approximately triangular shape and a circumference of around 0.50 m is captured from 3-dimensional human motion, and then the end-effector is allowed to track the trajectory within an error range of +/- 0.02 m. The simulation results suggest that the total tension of the four cables and the jerk value of the tracking trajectory both decrease with increasing training time. The proposed method will be validated later by actual experiments on the 3-DOF CDPR.