Deep reinforcement learning-based path following of a bio-inspired autonomous underwater vehicle under ocean currents

Bio-inspired autonomous underwater vehicles (BI-AUVs) offer high propulsive efficiency and maneuverability by imitating fish morphology and locomotion. However, oscillatory fin motion and fluid–structure interactions generate nonlinear hydrodynamic forces, making model-based controller design under environmental disturbances challenging. This study proposes a deep reinforcement learning (DRL)-based path-following framework for ocean currents conditions. A physics-based six-degree-of-freedom model incorporated ocean currents and the first-order response delay of fin actuators. A double-loop architecture combined a PID position controller with a DRL attitude controller. To mitigate simulation-to-real (Sim2Real) gap between the imperfect model and the physical platform, domain randomization varied mass, inertial parameters, and hydrodynamic coefficients. Curriculum learning progressively increased path complexity and current intensity to stabilize training. Compared with PD, NMPC, and baseline DRL controllers, the proposed controller reduced the cross-track error (CTE) RMSE by 83.0%, 47.1%, and 76.1%, respectively, in simulations with ocean currents. The trained policy was deployed directly on the platform and evaluated in a test basin under calm conditions. Although mission completion required 22.4 s longer than in simulation, the BI-AUV stably followed the desired path. These results demonstrate reduced sensitivity to model uncertainty, robust path following under currents, and successful transfer to the physical platform.