Reinforcement Learning-based Trajectory Planning for Cooperative Source Seeking

This paper investigates the cooperative source seeking problem via a networked multi-vehicle system, and proposes a reinforcement learning (RL)-based trajectory planning scheme for the system. In contrast to most existing works, the source position of this work is determined by multiple types of signal fields, and each vehicle in the network is responsible to take measurements of a type of signal field. In another word, a single vehicle cannot localize the source position and must cooperate with its neighbors. To cooperatively localize and simultaneously reach the source position, the vehicle is equipped with a trajectory planning scheme that combines consensus algorithm for cooperative source seeking and an RL-based algorithm for maximizing the value of its corresponding signal field. Thus, each trajectory planner only requires measurements of a field and relative positions between neighboring vehicles, which is especially appealing to global position system (GPS)-denied environments. Simulations are provided to validate the efficacy of the proposed trajectory planning scheme. The RL-based trajectory planner actively leads the vehicle to positions with higher field measurements, without the need for gradient or absolute position information. And we demonstrate the successful deployment of the model learned in single vehicle settings to the cooperative source seeking task.