Towards high precision control of strain wave gearing magnetic shape memory actuators utilizing reinforcement learning

This paper presents the first study to propose and evaluate control strategies for the recently introduced Strain Wave Gearing Magnetic Shape Memory (SWG-MSM) actuators, which are characterized by complex nonlinear dynamics of Magnetic Shape Memory Alloys (MSMA). Traditional model-based approaches are challenged by the lack of accurate and computationally efficient models for MSMA and consequently for SWG-MSM actuator class. Notably, despite the actuator's potential, no closed-loop position or velocity control strategies have been proposed for SWG-MSM actuators to date. As a baseline, a PID controller was tuned and further enhanced with an adaptive mechanism to compensate for changing actuator dynamics. To address the limitations of PID, a reinforcement learning (RL) approach using a Deep Q-Network (DQN) was developed and trained entirely in simulation. The RL agent was integrated into a cascade architecture with a proportional position controller, forming a P-DQN controller. The agent demonstrated strong generalization outperforming PID-based controllers across a wide range of reference velocities and external disturbances. Observation space optimization and output filtering further improved agents’ control accuracy, resulting in sub-micron positioning accuracy.