Reinforcement learning for robust navigation of fish-like agents in various fluid environments

Achieving robust and energy-efficient navigation in unknown fluid environments remains a key challenge for bioinspired underwater robots. In this study, we develop a reinforcement learning-based control framework that enables a fish-like swimmer to autonomously acquire effective navigation strategies within a high-fidelity computational fluid dynamics environment. By shaping the reward function to favor energy efficiency, the agent spontaneously discovers different locomotion patterns, ranging from continuous bursting to burst-and-coast gaits, all without prior knowledge of fluid mechanics. Although the agent is trained in a quiescent fluid environment, the learned swimming policies are generalized well in various navigation tasks and remain robust under complex flow perturbations, including uniform currents and unsteady vortex wakes. In all test scenarios, the agent achieves a 100%navigation success rate. These findings highlight the potential of integrating physics-based simulation with learning-based control strategy to advance the design of adaptive, efficient, and resilient aquatic robots inspired by biological swimmers.