Deep Reinforcement Learning Based Distributed Active Joint Localization and Target Tracking


We aim to solve a multi-robot active joint localization and target tracking (AJLATT) problem, where a team of robots with sensors of limited field of view cooperate with their neighbors to actively plan their individual motions so as to achieve better performance for self-localization and target tracking while avoiding collisions with their team members, the target, and the environment. To achieve that goal, a Deep Reinforcement Learning (DRL)-based distributed algorithm is proposed.Compared with other motion planning strategies, the DRL-based method has a near-optimal and long-sighted nature to solve the problem by learning from numerous trial-and-error interactions with the environment. Several simulations in different scenarios are performed to demonstrate the capability of our algorithm, and performance comparison with other motion strategies is given.

IEEE Robotics and Automation Society (ICRA) [Submitted]