Modular dual-critic reinforcement learning control for AUV path tracking and collision avoidance

Autonomous underwater vehicles (AUVs) operating in complex and uncertain ocean environments require reliable path-tracking and obstacle-avoidance capabilities to maintain safety in the presence of currents and obstacles. Traditional end-to-end deep reinforcement learning (DRL) methods often struggle with coupled navigation objectives, which can lead to oscillatory control and unreliable trajectory recovery. This paper proposes a modular dual-critic proximal policy optimization framework with prioritized experience replay (Modular DCPPO-PER) for AUV path tracking and collision avoidance. The framework separates navigation into two specialized policies for path tracking and collision avoidance and coordinates them through a time-to-collision-based soft risk arbitration mechanism. An actor-dual-critic-PER (ADCP) architecture is further introduced to improve learning stability and sample utilization by combining dual buffers with task-specific value estimation. In addition, a direction-sensitive reward function based on the sensor field of view assigns higher weights to high-risk perceptual regions and promotes smoother avoidance maneuvers. Simulation results in complex marine environments show that, compared with the advanced DRL baseline ARAB-PPO, the proposed method improves the average obstacle-avoidance success rate by 49.8% and reduces the path-tracking error by 12.9%. A USV-based physical surrogate experiment is also conducted to examine real-time inference, control continuity, and actuator-level feasibility on real hardware. The results indicate that the proposed framework can provide robust and smooth navigation behavior for AUV-oriented marine robotic applications, while full underwater AUV field validation remains future work.