Reinforcement Learning-Based Motion Control of Jellyfish Millirobots

Millirobots hold great potential for cutting-edge applications in targeted drug delivery and non-invasive diagnostics. Among them, the magnetic soft flapping wing millirobots (e.g., jellyfish-inspired millirobots) demonstrate excellent flexibility, safety, and adaptability across a wide range of Reynolds number regimes. However, due to the highly nonlinear nature of its environmental interaction model, motion control of the jellyfish millirobots remains a significant challenge. In this work, we employ a learning-based control framework and achieve model-free motion control for a 3 mm-long jellyfish millirobot through reinforcement learning and a Sim-to-Real approach. Additionally, we implement a path planning pipeline within this framework. The system achieves a 98% success rate in obstacle avoidance and a maximum positioning error of 0.26 body lengths during target hovering. The framework's model-free nature provides strong generalizability for controlling diverse magnetic millirobotic systems, overcoming traditional limitations imposed by complex physics-based modeling and significantly enhancing operational accessibility for various applications. The project page is https://zyl-hub.github.io/multimodal-jellyfish-microrobot .