Model-Based Reinforcement Learning for Position Control of Continuum Manipulators Actuated by Pneumatic Artificial Muscles

The topic of position control for continuum manipulators (CMs) remains open and yet to be well explored and developed. Current applications of CMs focus on employing static or quasi-dynamic controllers built upon kinematic models or linearity in the joint space, resulting in a loss of the rich dynamics of a system. This paper presents a model-based reinforcement learning scheme for position control of a class of CMs with strong nonlinearity and input coupling, which includes a probabilistic dynamics model as the dynamic forward model and a policy update approach for the closed-loop policy. The effectiveness of the scheme is verified on a dual-segment CM actuated by pneumatic artificial muscles, and the experimental results confirm that such scheme can obtain good results with only a limited number of samples and interactions.