Performance evaluation and optimization of a mobile robot path-planning model based on reinforcement learning

Abstract

Autonomous navigation in mobile robotics faces significant challenges in dynamic, unstructured environments, as re-planning strategies often lack time efficiency when recomputing around new obstacles. Traditional path-planning algorithms typically rely on static, pre-mapped environments and rigid geometric rules, which severely limit their adaptability when unexpected obstacles appear. Reinforcement learning provides a powerful, data-driven solution to these limitations by enabling autonomous agents to continuously learn and refine optimal navigation policies through trial-and-error interactions, prioritizing the most rewarding paths. It is with this in mind that this work proposes using reinforcement learning, specifically Q-learning, for a mobile robot path-planning model, global dynamic value iteration for re-planning when unexpected obstacles are detected, and evaluating and optimizing multiple model variables to improve both efficiency and correctness. Experimental results include implementing and executing the proposed model on a mobile robot platform, as well as using a 3D LiDAR depth sensor to detect and evaluate the environment. Key areas, such as dynamic reward shaping, active perception, and safe path detection, are addressed to overcome the inherent limitations of model-free learning. The expected result is an intelligent system that autonomously navigates a dynamic environment with robust precision, with potential applications in industrial automation, logistics, and healthcare.