Adaptive Actuation of Magnetic Soft Robots Using Deep Reinforcement Learning

Abstract

Magnetic soft robots (MSRs) have attracted growing interest due to their unique advantages in untethered actuation and excellent controllability. However, actuation strategies of these robots have long been designed out of heuristics. Herein, it is aimed to develop an intelligent method to solve the inverse problem of finding workable magnetic fields for the actuation of strip‐like soft robots entirely based on deep reinforcement learning algorithms. Magnetic torques and a dissipation force to the Cosserat rod model are introduced, and the developed model to simulate the dynamics of MSRs is utilized. Meanwhile, under the reinforcement learning framework, soft robots to move forward without human guidance are successfully trained, and the results intelligently adapt to different magnetization patterns and magnetic field restrictions. The learned actuation strategies by directly applying simulated magnetic fields to real MSRs in an open loop way are validated. The experimental results show good accordance with simulations. By presenting the first case of using strategies entirely generated by reinforcement learning to control real MSRs, the potential of using reinforcement learning to achieve autonomous actuation of MSRs is demonstrated, which can be used to establish a route for the creation of highly adaptive design framework.

Keywords

• Cosserat rod model
• deep reinforcement learning
• magnetic actuation
• magnetic soft robotics
• magnetization