CS289 (Biologically Inspired Multiagent Computing) Final Project

Developing Multi-Agent Approaches for Active Shape Detection and Analysis

Many thanks for Michael Kester for being a wonderful partner! The project code can be found here and the report can be downloaded here.

Summary: The inspiration for this project was the idea of being able to scout an unknown environment and determine safe routes that more advanced (and expensive) robots could take. One approach that I’ve worked on before was using computer vision to identify roads from a UAV aerial feed. This project takes an approach of using swarm robotics. The question is whether we could use a swarm of robots that maintain a fixed formation to determine how far they were going. It turns out that maintaining a fixed formation moving in a straight line without any type of global locating service (like GPS) is not easy at all. Indeed, most current multi-agent systems have an overhead camera service that can precisely track where each robot is at a given time, which is infeasible in the long-run. In this projects, we use Kilobots, a low-cost swarm robot that can scale up to over a thousand robots in a single group but can only sense the distance between one another. We wanted to see how far we could push the envelope with just this one feature.

Abstract: The identification of traversable terrain within an environment is essential for the development of autonomous robotic networks capable of operating over a wide area within a rapidly evolving environment. We propose the usage of many robots with fixed formations to be able to move through an environment and provide the necessary odometry to map out the region. We design and implement an algorithm for a group of four Kilobots, robots that are capable of broadcasting communication, motion, distance sensing, but not odometry to walk together in a straight line with a square formation. Our algorithm is able to provide consistent motion that far outstrips the capability of a single robot. Specifically, we find that algorithms that try to fix the angles between constituent robots perform well in maintaining a prescribed bearing, but tend to drift in the long run. However, algorithms that try to move first to keep their lateral edges parallel take a while to equilibriate, but tend to hold their bearing and formation for longer without any type of global error correction. We hope that our studies can serve as a springboard for future work on odometry with formation control.