Self-learning how to swim at low Reynolds number

Abstract

Machine learning is integrated into low-Reynolds-number locomotion to enable a class of self-learning, adaptive (smart), micro-swimmers. Instead of specifying locomotory gaits in advance, a self-learning swimmer develops and adapts its propulsion strategy based on interactions with the surroundings via reinforcement learning. Without requiring prior knowledge, the swimmer can recover previously known propulsion strategies, and improve and adapt in different media. This development can enable the design of smart micro-robots with robust locomotive capabilities in complex environments