Research on 3D Obstacle Avoidance of Autonomous Underwater Vehicle Based on Deep Reinforcement Learning

Formation control and obstacle avoidance are critical challenges in the field of underwater robotics. This paper presents a novel approach using multi-agent reinforcement learning to address the critical challenges of formation control and obstacle avoidance for underactuated autonomous underwater vehicles (AUVs) operating in three-dimensional space. The proposed framework implements a leader-follower scheme, wherein the leader guides the team toward a designated target location while the followers maintain formation and avoid obstacles. To enhance training efficiency, the state space representation of the AUVs is defined in the body coordinate system, filtering irrelevant information. Extensive simulations validate the efficacy of the proposed approach, demonstrating its capacity to achieve formation control and obstacle avoidance with high adaptability and efficiency in complex underwater environments. The results highlight the potential of multi-agent reinforcement learning techniques to significantly advance the capabilities of underwater robotics systems.