Behaviour-Aware Adaptive Path Planning With Q-Learning Weight Adaptation

Abstract

This article proposes a dynamic path cost modelling framework that integrates fuzzy C-means clustering, association rule mining, and a Q-learning-based weight adaptation mechanism to achieve multi-objective collaborative path planning based on active perception of obstacle behaviour patterns. The method constructs an obstacle behaviour pattern library from multimodal perception, mines context-dependent mappings between environment features and the weights of three cost terms, namely energy, safety margin, and path smoothness, and uses Q-learning to adjust these weights online. The planner is embedded in a two-layer navigation scheme, where an improved artificial potential field, APF, generates a global reference path and a dynamic window approach, DWA, local planner executes it under dynamic obstacles. Experiments on the ROS/Gazebo platform with TurtleBot3 in environments containing up to 25 obstacles per 100 m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> show that the proposed adaptive planner maintains a success rate of about 80% in high-density scenes, improving over classical APF by 30% and DWA by 55%, and remaining comparable to the deep reinforcement learning baseline TD3 at 82%. At the highest obstacle density, the method reduces normalised energy per metre by 9.4% and 6.1% compared with APF and DWA, respectively, while improving the smoothness index by 6.7–12.7% and the safety-margin index by 25–42.9%. These results indicate that the framework effectively balances real-time performance, energy efficiency, and safety, and provides a promising building block for adaptive multi-robot collaboration under communication constraints.