Stochastic Potential Game-Based Target Tracking and Encirclement Approach for Multiple Unmanned Aerial Vehicles System

Le résumé fourni par la source

Utilizing fully distributed intelligent control algorithms has enabled the gradual adoption of the multiple unmanned aerial vehicles system for executing Target Tracking and Encirclement missions in industrial and civil applications. Restricted by the evasion behavior of the target, current studies focus on constructing zero-sum game settings, and existing strategy solvers that accommodate continuous state-action spaces have exhibited only modest performance. To tackle the challenges mentioned above, we devise a Stochastic Potential Game framework to model the mission scenario while considering the environment’s limited observability. Furthermore, a multi-agent reinforcement learning method is proposed to estimate the near Nash Equilibrium strategy in the above game scenario, which utilizes time-serial relative kinematic information and obstacle observation. In addition, considering collision avoidance and cooperative tracking, several techniques, such as novel reward functions and recurrent network structures, are presented to optimize the training process. The results of numerical simulations demonstrate that the proposed method exhibits superior search capability for Nash strategies. Moreover, through dynamic virtual experiments conducted with speed and attitude controllers, it has been shown that well-trained actors can effectively act as practical navigators for the real-time swarm control.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.