MAP3F: a decentralized approach to multi-agent pathfinding and collision avoidance with scalable 1D, 2D, and 3D feature fusion

Path planning and collision avoidance are vital aspects of successful development and utilization of robots in complex and multi-agent environments. With the integration of robots into social settings, the significance of this issue becomes more apparent. This paper introduces a decentralized management approach based on deep reinforcement learning, where each agent learns independently based on its local observations. The proposed method employs a feature fusion technique which combines 1D, 2D, and 3D features. In order to streamline computation and optimize the training process, an established separation index method is utilized. This approach strategically selects a subset of the most informative features. The presented approach outperforms classical and learning-based methods in various environments with differing densities. Performance evaluation metrics include the interaction index, which indicates the percentage of collision-free scenarios, the reachability index, measuring the time for the slowest agent to reach its goal, the field of view index, demonstrating reduced computation time by narrowing the field of view without compromising interaction, and the scalability index, quantitatively measuring a system’s capability to efficiently handle increasing amounts of work or its ability to be enlarged to accommodate that growth. The performance of this method, compared to PRIMAL, ORCA, and ODRM* methods, has shown an increase of over 30% in situations where the environment is more complex and the number of agents is higher.