A Deep Reinforcement Learning-Based Tracking Method for Monitoring Tasks with Improved Obstacle Avoidance and Training Strategy

In response to the challenges posed by an aging population and the lack of child guardians, the development of guardian robots has become increasingly critical. A key challenge is effectively tracking targets in complex environments, which are characterized by unpredictable movements and the presence of both dynamic and static obstacles. This paper proposes an advanced tracking method based on deep reinforcement learning to address these challenges. Specifically, we introduce an action mask to constrain the robot's movements, enhancing its dynamic obstacle avoidance capability. Additionally, we employ a multistage training strategy inspired by curriculum learning, which streamlines the training process and improves convergence. By overcoming limitations of traditional methods, our approach adapts efficiently to dynamic environments. Experimental results demonstrate that the proposed method not only achieves higher tracking success rate and sustained tracking capability, but also significantly improves the robot's ability to generalize to various targets and respond dynamically to complex environments. These results underscore the effectiveness and potential of our approach for real-world applications in elderly care and other monitoring tasks.