Development of a multi-robot testbed for direct human-swarm interaction
Committee members: Fonseca, Benedito; Ryu, Ji-Chul.||Advisor: Butail, Sachit.||Includes bibliographical references.
Abstract
Robotic swarms are well suited for missions that involve covering large areas such as environmental monitoring, search-and-rescue operations, as well as those that require fusing information from multiple sources such as target tracking and machine inspection. However, navigating a swarm of robots effectively requires performing higher-level functions that are typically executed by humans such as commanding them to stay close and guiding them along with a preplanned route. In this context, the ability of multiple robots to stay together autonomously significantly reduces the cognitive load on a human operator. The goal of this thesis is to design and build a multi-robot testbed for studying human-swarm interaction. The specific objective of this research is to use methods from simulation-based design to estimate the control parameters required for robust operation of the robots. A consensus algorithm is implemented on simulated unicycle agents to keep them within sight of each other, as they follow a human leader. The algorithm is augmented with features to match a real-world implementation including realistic occlusion of swarm members in the presence of obstacles and collision handling. Search strategies are incorporated to reduce instances where a robot can be lost forever. Following an exhaustive simulation exercise, the control gains that reveal the best performance in terms of swarm connectedness and the ability to reach a goal location are further assessed in terms of their sensitivity to sensor noise. Finally, the approach is implemented and tested on an iRobot Create 2 testbed that consists of six robots monitored using an array of wide-angle web cameras in a large laboratory environment with obstacles. In a true decentralized architecture, the robots are instrumented with a microcontroller and a web camera to follow each other using vision-based cues only without any external reference.
