Abstract
This paper proposed an optimized navigation method based on Deep Deterministic Policy Gradient (DDPG) to address the issues of poor navigation paths ,and resolved the robot system encountered the blocked exits in narrow areas during collaborative transportation. The method aimed at generating optimized navigation paths for leader-follower robot formations to complete tasks such as moving objects, designing swap decision reward function for DDPG to resolving the problem of blocked exits in narrow areas. Specifically, the paper first optimizes the reward function module in the DDPG network to incorporate a decision-swapping reward mechanism for training the formation's navigation capability. Next, it utilizes Unscented Kalman Filter (UKF) filtering to estimate the formation's position states for follower trajectory tracking. Finally, the navigation performance of the formation is validated through simulations. The results of the simulation experiments demonstrate that the formation can achieve a navigation success rate of around 95%, in random environments. Additionally, compared to paths generated by A*-DWA and RRT* algorithms, the trained DDPG navigation algorithm reduced the average angle variation of generated paths by 64.03% and 38.65%, respectively. Furthermore, in trap environments, the formation is capable of executing decision swapping to exit restricted areas and generate a passable path for the formation.