A Survey on Collision Avoidance Algorithms for Multi-robot Systems

Multi-robot systems enable cooperation between multiple robots to achieve common goals or tasks. These systems can enhance efficiency and productivity in various applications, such as transportation, manufacturing, and exploration. However, a critical issue in multi-robot systems is the possibility of collisions with both static and dynamic obstacles. This survey presents the latest trends and advancements in collision avoidance approaches for multi-robot systems. We analyze classical methods focusing on centralized and decentralized collision avoidance. We categorize and examine decentralized methods based on several criteria as velocity obstacles, communication, and control, etc. In addition to classical methods, we explore learning-based methods for collision avoidance in multi-agent systems. We categorize and analyze these methods into two types: Reinforcement Learning-based (RL) methods and large language model-based (LLM) methods. Within RL-based methods, we identify three main approaches: 1) Purely reward-based approaches, which rely solely on rewards to avoid collisions, 2) Approaches based on mixing reward-based and training guidance, 3) Approaches based on mixing reward-based and classical control methods. LLM-based methods operate independently of reward-based approaches, using language-based decision-making to generate trajectories.