A Hierarchical Tracking Framework With Adaptive Skill Utilization and Experience Sharing

Abstract

Active tracking technology holds critical applications in dynamic scenarios ranging from autonomous driving to service robotics. Therefore, developing efficient and robust tracking strategies is crucial. While deep reinforcement learning has emerged as a promising approach for complex tracking demands, most existing methods face significant challenges when handling intricate tasks: manually designed dense reward functions and complex network architectures not only induce overfitting risks but also entail substantial labor and temporal costs. To address this, we propose a hierarchical reinforcement learning based tracking framework, whose core innovation lies in decoupling low-level task skills from high-level decision-making logic through task decomposition. The selection agent adaptively selects task skills, then activates corresponding modules of the capability agent to output actions. This achieves model simplification while maintaining performance and enhancing interpretability. To improve sample utilization efficiency, we propose a cross task experience sharing mechanism that enables skill synergy through shared training samples, further boosting model performance. Experimental results demonstrate that this mechanism effectively enhances the performance of all task skills in the capability agent while ensuring stability. Simulation experiments reveal that compared to the baseline models, our framework exhibits superior task performance and interpretability through adaptive skill switching.