Multi-Agent Path Finding (MAPF) in dynamic environments is challenging due to limited observations, non-cooperative moving obstacles, and inefficient inter-agent coordination. Existing search-based and reinforcement learning methods often struggle to balance real-time adaptability, collision avoidance, and scalable cooperation. To address these issues, this paper proposes MA-DyRoLT, a MAPPO-based MAPF framework that integrates dynamic waypoint generation and learned communication topology. The dynamic waypoint mechanism provides adaptive local guidance according to environmental openness, while the learned topology enables sparse and task-relevant information sharing among agents. In addition, a Transformer-based temporal encoder is introduced to capture historical observation dependencies, and an auxiliary pedestrian trajectory prediction module with a prediction-based risk penalty is designed to encourage proactive collision avoidance. Simulation results demonstrate that MA-DyRoLT outperforms representative search-based and learning-based baselines in success rate, movement time, and runtime. It maintains high success rates in large-scale dynamic scenarios, achieving 0.998 success rate in a 120 × 130 environment with 150 agents. These results verify the effectiveness, scalability, and real-time potential of the proposed method.
Liang et al. (Wed,) studied this question.