End-to-end deep reinforcement learning enables bionic fish to learn efficient obstacle avoidance in unknown underwater environments

Path planning is a core issue in fields such as robotics, autonomous driving, and unmanned navigation systems. Nevertheless, traditional path planning methods face substantial challenges in dynamic, nonlinear, and stochastic underwater environments. In this study, a fluid solver module capable of simulating fluid–structure interaction with high precision and efficiency was constructed based on the immersed boundary–lattice Boltzmann method. The perception and decision-making system of the bionic intelligent fish was constructed using deep reinforcement learning (DRL) algorithms, and a concise visual perception modality representation based on one-hot encoding was proposed. The obstacle avoidance performance of the bionic intelligent fish trained via end-to-end DRL was tested in both still-water environments and dynamic flow environments with complex geometries. The results demonstrate that the intelligent fish exhibits efficient local path planning and obstacle avoidance capabilities in both environments while possessing a certain robustness against disturbances in nonlinear environments. This research can provide a reference for the design of local path planning modules for bionic robotic fish and is expected to further enhance the mission execution capabilities of autonomous underwater vehicles in unmapped deep and far-sea areas with high uncertainty.