Control of Continuum Manipulators with Shape Constraints via Deep Reinforcement Learning

Abstract

Continuum robots, while versatile for handling complex tasks, present significant challenges in control system design. This paper introduces a novel framework that integrates position and orientation control laws to address shape constraints effectively. Specifically, a Deep Reinforcement Learning (DRL) strategy is proposed to facilitate trajectory tracking within the desired orientation. To ensure safety, the framework directly controls the end-effector's position and orientation in the workspace, avoiding unsafe zones. A centralized control law, the Modified Transpose Jacobian (MTJ), is employed to resolve the robot's redundancy without relying on inverse kinematics. The proposed approach is validated through simulations on a Tendon-Driven Continuum Robot (TDCR), demonstrating superior performance compared to similar learning-based controllers.