A Framework for Controlling Robotic Swarms Using Bayesian Optimization and Linear Combination of Vectors

Stephen Jacobs*, R. Michael Butts*, David Rubel*, Yu Gu, Ali Baheri, and Guilherme Pereira
Statler College of Engineering and Mineral Resources, West Virginia University, Morgantown, WV 26505

Presentation No.: 59

Assigned Category (Presentation Format): Engineering (Poster Presentations)

Student’s Major: Mechanical and Aerospace Engineering

Interest in robotic swarm research has grown in recent years. In contrast to single-agent or multi-agent systems which often use complex centralized control schemes, robotic swarm systems often rely on local sensing and decentralized control schemes to accomplish complex tasks through emergent behaviors. In these systems, it is often very challenging to predict the outcomes of even the simplest control rules. We propose a generalizable control and optimization framework for decentralized direct-communicating robotic swarms. The proposed framework facilitates the discovery of control laws leading to emergent behaviors by simplifying the control policy to a linear combination of vectors and scalars which are collected from sensor inputs. The weights for each term in the control function are found by simulating the swarm system and optimizing a given cost function using Bayesian Optimization. Using this technique we are able to generate emergent behavior to suit a number of swarm tasks such as cohesion and segregation, and collision avoidance.

Funding:

Program/mechanism supporting research/creative efforts: the WVU Robotics REU (Yu Gu & Jason Gross)