Multi-robot Collision Avoidance in Non-communication Deep-Sea Environments via Reinforcement Learning

This is the authors' abstract. We don't add key points for this paper.

Ensuring effective multi-robot collision avoidance in deep-sea environments presents unique challenges due to the absence of direct communication and the intermittency of positional updates. Existing methods struggle with outdated information, leading to suboptimal decision-making and increased risk of collisions. In this paper, we address these issues by introducing a multi-agent reinforcement learning (MARL)-based framework tailored for multi-robot navigation under intermittent observations. Our approach incorporates a recurrent neural network (RNN)-based module to take advantage of historical information, enabling agents to adapt to dynamic environments despite communication constraints. By leveraging MARL, our method effectively balances safety, efficiency, and path optimality, allowing robots to make informed decisions even with limited sensing updates. Simulations demonstrate that our approach significantly outperforms existing baselines, achieving higher success rates and improved adaptability in high-density, long-interval scenarios.