Learning to avoid: decentralized multi-robot navigation from LiDAR using PPO and reciprocal velocity obstacles

Abstract

Autonomous navigation in dynamic multi-agent environments remains a fundamental challenge in robotics, particularly when real-time coordination and collision avoidance are required without communication. This paper proposes a decentralized, model-free, multi-robot navigation framework that leverages LiDAR-based perception and the Reciprocal Velocity Obstacle (RVO) formulation within a deep reinforcement learning (DRL) paradigm. The system uses Bi-GRU architectures to process variable-length exteroceptive observations and applies the Proximal Policy Optimization (PPO) algorithm to learn navigation policies. A custom simulator with performance enhancements was developed to efficiently model dynamic and static obstacles using simulated LiDAR data. The model demonstrated strong performance in static and dynamic environments with varying agent counts, achieving high success rates and adaptive behaviors. By combining geometric motion planning with modern DRL methods, the framework achieves robust, scalable, and communication-free navigation, offering a promising direction for future multi-robot systems in real-world applications.