Decentralized Reinforcement Learning for Cooperative Multi-Robot Navigation

Cooperative multi-robot navigation requires large teams of robots to reach individual goals without collisions in shared spaces, which is a challenge compounded by partial observability, dynamic interactions, and the combinatorial nature of joint decision-making. Centralized methods provide optimality but fail to scale, while decentralized approaches offer efficiency yet often suffer from deadlocks and weak cooperation. To overcome these limitations, we propose a decentralized reinforcement learning framework that integrates a dual-head soft actor-critic architecture with selective graph-attention communication and task-specific reward shaping. A compact encoder processes a novel observation representation that combines local fields of view, goal projections, and directional heuristics, enabling scalable awareness under limited sensing. Heuristic and blocking rewards further accelerate convergence and promote cooperative navigation in congested regions. Extensive experiments on grid environments of varying sizes and robot densities show that our method achieves faster convergence, higher success rates, and stronger scalability than state-of-the-art methods in this domain.