Learning-Based Cooperative Mobility Control for Autonomous Drone-Delivery
Soohyun Park, Chanyoung Park, Joongheon Kim
Abstract
It has been widely considered that autonomous aerial drone-delivery will play a key role in next-generation logistics. For enabling robust and reliable autonomous multi-drone aerial package delivery services, the proposed algorithm has to mainly consider two objectives, i.e., (i) the maximization of total number of delivered packages and (ii) the efficient energy utilization which is defined as the energy consumption facilitation for increasing the number of delivered packages while avoiding battery exhaustion. This can be formulated as a kind of scheduling optimization problems, after that, this is re-formulated as a multi-agent deep reinforcement learning (MADRL), in order to convert integer programming-based formulation which is hard to solve in polynomial-time into MADRL-based discrete-time sequential decision making over multiple cooperative drones. For the solution approach of MADRL, a novel communication network (CommNet)-based algorithm is designed for multi-drone cooperation in order to optimally achieve our two objectives. Our performance evaluation results verify that the proposed CommNet-based algorithm achieves desired performance improvements in terms of the maximization of total number of delivered packages as well as the efficient energy utilization.