Reinforcement Learning on Modular Robots Using Real-Time Decentralized Deconfliction in a Warehouse Environment

Abstract

This work seeks to incorporate existing methodologies for controlling and localizing robotic hardware of different capabilities in order to create an autonomous system of robots to move deliverables within an established environment. The system utilizes modular designs to employ rapid prototyping of different robotic capabilities highlighting the adaptability of this methodology. A high-level task manager centrally assigns priorities to each deliverable, and, through decentralized planning, the robots deconflict with one another during execution. Through the adoption of reinforcement learning techniques, our robots create individual path plans that adapt to both static and dynamic obstacles in real time. Contrasting other works that manage robotic systems in a predetermined and fixed environment, we demonstrate a single, simplified approach that can control and localize modular robots simultaneously in any given setting, to include warehouses and factories. This work utilizes the Robotic Operating System (ROS) to communicate between multiple robots of various capabilities and a single task manager, synchronizing autonomous movement and avoiding other tracked objects.