DS-MRNav: Distributed Multi-Robot Navigation With Static Obstacles via a Decoupled-State Dual-Stream Framework

Abstract

Multi-robot cooperative navigation is important in various applications, such as warehouse automation and large-scale manufacturing. However, the coupling between dynamic interactions and static environmental constraints often causes deadlock and loss of efficiency in decentralized multi-robot navigation. To address this issue, this letter proposes a decoupled-state dual-stream framework for distributed multi-robot navigation in static-obstacle environments (DS-MRNav). Different from existing methods that mix heterogeneous observations directly, the proposed approach decouples dynamic interaction constraints and static obstacle constraints at the state level, and adopts a dual-stream architecture to process these two types of local perception. The dynamic and static streams are each fused with the ego-state, and the resulting features are further fused to generate cooperative actions, while the policy is trained via curriculum learning and a safety-aware hybrid reward. In simulation, DS-MRNav shows strong robot-density scalability: in the empty-scenario scalability test, it generalizes to 20 robots and achieves a 99% episode success rate. On static-obstacle maps, it maintains high robot-level success and low timeout rates on both in-distribution and unseen maps. Ablation studies validate the contributions of the decoupled architecture, ORCA half-plane representation, safety-aware reward terms, and frozen pretrained static encoder. Real-robot experiments further show stable cooperative avoidance and obstacle circumnavigation without additional fine-tuning, indicating the promising sim-to-real transferability of the proposed approach.