Deep reinforcement learning and graph-based approaches for multi-robot collision control

Abstract

Recently, multi-robot systems, in which several robots coordinate along with each other to achieve clearly defined goals, have become increasingly popular. But the challenge is to schedule them efficiently to avoid collisions and execute their respective jobs in the least time possible. The simplicity of multi-robots led to a wide range of potential applications, and we will focus on applications in the process of fiber placement. Fiber placement refers to a fabrication process for composite materials where reinforcing fibers are placed along a predetermined path in the component. The present thesis proposes various approaches like reinforcement learning and graph-based methods for optimal collision control in a multi-robot system and to schedule the robots to execute their tasks in the least possible time. These can be compared and later employed to significantly reduce the lead time in the multi-robot fiber placement process.