Multi-Agent Reinforcement Learning for Coordinated Motion: Integrating Potential Field based Rewards

Multi-agent coordination for autonomous navigation faces challenges due to sparse reward signals and coordination deadlocks when agents learn independently. This thesis presents a multi-agent reinforcement learning framework that integrates artificial potential field principles with adaptive role assignment for two-agent coordinated navigation. The approach introduces a reward mechanism that dynamically assigns roles based on agents' relative distances to the goal. This mechanism addresses coordination deadlocks that occur when agents receive conflicting directions from symmetric potential field calculations. The framework uses Multi-Agent Proximal Policy Optimization (MAPPO) with centralized training and decentralized execution. Agents learn policies based on local LiDAR observations and relative position information, without requiring communication during execution. Experimental evaluation compared three reward configurations across 250 standardized episodes with challenging obstacle configurations featuring 120 obstacle cells distributed in realistic cluster patterns. The proposed approach achieved a 63.2% success rate, compared to 50.0% for potential field guidance alone and 44.4% for a baseline collective progress reward. Analysis of learned behaviors reveals that agents develop adaptive coordination strategies that enable effective navigation in constrained environments. While limited to two-agent scenarios in 2D simulation, the approach provides a basis for studying physics-informed reward design in multi-agent coordination tasks.