Deep Reinforcement Learning for Multi-Robot Local Path Planning in Dynamic Environments

The transportation and logistics sectors are experiencing a remarkable increase in the utilization of autonomous mobile robots, playing a pivotal role in the efficient management and distribution of merchandise and resources. Particularly in large-scale systems, finding effective solutions for coordinating and planning paths is a critical factor influencing overall system performance. Addressing the challenges of dynamic environments and working in uncertain conditions, this study introduces a novel local path planning approach for a multi-agent system based on a decentralized framework, employing deep reinforcement learning. The methodology incorporates Proximal Policy Optimization (PPO) and the Dynamic Window Approach (DWA) to analyze situations based on environmental information. We propose a new cost function for DWA by developing two subfunctions based on the information between agents and the goal. The observation space for deep reinforcement learning is designed by integrating velocity space and incorporating the cost function derived from DWA. The effectiveness of this method is evaluated in four simulated environments and experimented with four TurtleBot Burger robots. Additionally, the approach is compared with state-of-the-art multi-robot path planning methods, revealing significant improvements in success rates, reaching up to 30%, in dynamic environments, and reductions in path lengths.