RL-Inspired Pairwise Path Selection for Multi-Agent Navigation

Abstract

This paper presents an RL (Reinforcement Learning) inspired pairwise path selection framework for multi-agent navigation in an offline setting. Path performance is quantified using interaction metrics derived from artificial potential field (APF)-based collision avoidance, together with overall task completion time. These metrics are incorporated into a reward formulation that captures both safety and efficiency. A neural network model is trained using pairwise comparisons of path executions, enabling the system to learn relative preferences among candidate path. This formulation eliminates the need for continuous environment interaction during training while preserving key characteristics of reward-driven optimization. The framework is validated through simulations implemented in ROS 2 (Robot Operating System) using TurtleBot4 platforms in a warehouse environment. Multiple path combinations are evaluated under varying interaction conditions. Results demonstrate stable training convergence, high ranking accuracy, and clear discrimination between efficient and conflict-prone paths.