- This event has passed.
ESE Thesis Defense: “Scalable Learning in Distributed Robot Teams”
April 21 at 11:00 AM - 1:00 PM
Mobile robots are already in use for mapping, agriculture, entertainment, and the delivery of goods and people. As robotic systems continue to become more affordable, large numbers of mobile robots may be deployed concurrently to accomplish tasks faster and more efficiently. Practical deployments of very large teams will require scalable algorithms to enable the distributed cooperation of autonomous agents. We focus on the three main algorithmic obstacles to the scalability of robot teams: coordination, control, and communication.
To address these challenges, we design graph-based abstractions that allow us to apply Graph Neural Networks (GNNs). First, a team of robots must continually coordinate to divide up mission requirements among all agents. We focus on the case studies of exploration and coverage to develop a spatial GNN controller that can coordinate a team of tens of agents as they visit thousands of landmarks. A routing problem of this size is intractable for existing optimization-based approaches.Second, a robot in a team must be able to execute the trajectory that will accomplish its given sub-task. In large teams with high densities of robots, planning and execution of safe, collision-free trajectories may require the joint optimization over all agent trajectories, which is impractical in large teams. We present a controller for the problem of flocking that uses delayed communication formalized via a GNN to allow aerial robots to avoid collisions and align velocities. Third, robot teams may need to operate in harsh environments without existing communication infrastructure, requiring the formation of ad-hoc networks to exchange information. Many algorithms for control of multi-robot teams operate under the assumption that low-latency, global state information necessary to coordinate agent actions can readily be disseminated among the team. Our approach leverages GNNs to control the connectivity within the ad-hoc network and to provide the data distribution infrastructure necessary for countless multi-robot algorithms.
ESE Ph.D. Candidate
Ekaterina (Kate) Tolstaya received her B.S. in Electrical Engineering and Computer Science from the University of Maryland, College Park, in 2016. She received her M.S. in Robotics from the University of Pennsylvania in 2017. Her interests lie at the intersection of robotics and machine learning, with a particular emphasis on multi-robot and multi-agent systems. Kate is grateful for the support of the National Science Foundation as a Graduate Research Fellow. Kate was an intern at Microsoft Research, DeepMind, and Waymo, and she will be returning to Waymo as a Research Scientist after the completion of the doctorate program.