Comparative Analysis of Classic and Reinforcement Learning Approaches for Robot Navigation in Dynamic Environments

Path-finding is a fundamental problem and crucial in robotics. It enables robots to navigate the environment and complete their assigned tasks efficiently. Classic path-finding solutions such as A* typically use re-planning tactics when faced with dynamic impediments. These strategies involve calling back a re-planning algorithm to find a different path each time the robot encounters a dispute. However, these techniques frequently require intensive and extra processing and computation. We propose a distributed solution based on multi-agent reinforcement learning to overcome this issue. The proposed method considers two aspects: the global aspect for a static environment and the local aspect applied in a dynamic environment, which is divided into sub-environments, each assigned to an agent. Each agent uses reinforcement learning for local path planning in response to changes, enhancing scalability and reducing computational demands. Additionally, the performance is improved by applying the bankruptcy concept. Simulation results indicate that the proposed method has promising potential to be applied to path-finding problems.