Autonomous underwater vehicle obstacle avoidance using deep reinforcement learning

This study presents a deep reinforcement learning framework integrating convolutional neural networks for autonomous underwater vehicle navigation and evasion. Sonar signals are converted into spatial representations through CNN-based feature extraction and then processed by a deep Q-network to support real-time decision making in dynamic, noisy underwater environments. Experience replay and target networks improve training stability, convergence speed, and adaptability to environmental changes. The proposed method enhances perception, action selection, obstacle avoidance, and path optimization compared with traditional approaches. By combining convolutional spatial learning with value-based decision modeling, the framework improves generalization, safety, efficiency, and operational autonomy of AUVs operating in complexity.