Learning-Based Motion Controller for Reconfigurable Microswarms

Motion control of magnetic microswarms has attracted extensive attention due to its significance in microrobots-based biomedical applications such as targeted drug delivery. However, such reconfigurable microswarms are subject to complex interactions between individuals and environments which make accurate modeling challenging. These complexities of microswarms poses challenges for precise motion control, as traditional controllers often rely on precise mathematical models and manual parameter tuning that limits their scalability and efficiency. Learning-based methods, such as Deep Reinforcement Learning (DRL), offer an alternative but require large datasets (usually on the order of millions) and extensive exploration which may cause the microswarms instability in physical environments due to unreasonable actions during early training therefore results in the sim-to-real gap. Moreover, traditional DRL focuses on instantaneous state-action mappings, neglecting the sequential dependencies critical for accurate motion control, leading to low tracking accuracy in complex scenarios. To address these challenges, we propose a Learning from Demonstration (LfD)-based motion control framework, which inherently encode compensatory behaviors and task-specific adaptability into neural networks, enabling adaptive performance even under unmodeled disturbances. Furthermore, the neural networks consider a time series of microswarm states to determine the future control actions, enabling the system to learn sequential dependencies and transitions between states so as to ensure smooth and accurate motion control. Simulations and comparative experiments validate our framework’s effectiveness and demonstrate superior control accuracy and adaptability to microswarm’s shape changes.