Learning to Plan Quickly is Key to Adaptive and Efficient Multi -Robot Operations & Urban Air Mobility Networks

Adaptation is fundamental to the performance of multi-robotic systems and cyber-physical networks that operate in complex uncertain environments. For example, in multi-robot or swarm systems that provide task parallelism, resilience, and inexpensive deployment options, adaptation demands the ability to offer consistent performance across different-sized task spaces, and various environment scenarios that often involve natural and adversarial uncertainties. Time-sensitive planning of agent activity is key to such adaptation. In this context, the seminar talk will touch upon our work in three major areas: 1) Time-critical operations such as disaster response and reconnaissance where teams of robots must coordinate to complete tasks under environment variations and adversities. 2) Fleet planning and vertiport air-traffic-management in urban air mobility (aka flying taxi) networks. 3) Reconfiguration of disrupted power grid networks. The majority of the above problems can be modeled as combinatorial optimization (CO) problems, exact solutions to which are however too expensive (usually prohibitive) to achieve within the necessary planning time frames. In recent years, an exciting body of work has emerged on using special reinforcement learning-based techniques to model approximate solutions or intelligent heuristics for such (CO) problems. Motivated thereof, we have innovated a few learning frameworks and compatible simulation environments to tackle the generalizability, scalability, and robustness needs better than state-of-the-art alternatives, while keeping training costs reasonable. Notable among these are a class of graph neural network architectures that leverage the concepts of capsule networks, attention mechanisms, and topological descriptors of graphs; and a neuroevolution algorithm that evolves parsimonious neural policy models suited for tactical planning.

Dr. Souma Chowdhury is an Associate Professor of Mechanical and Aerospace Engineering, with adjunct affiliation in Computer Science and Engineering, at University at Buffalo. There he leads the Adaptive Design Algorithms, Models, and Systems (ADAMS) Lab, and serves as the Co-Director of the new UB Center for Embodied Autonomy and Robotics (CEAR). Dr. Chowdhury graduated with his Ph.D. in Mechanical Engineering from Rensselaer Polytechnic Institute (RPI) in Troy, NY. Prior to joining U Buffalo, he worked as a Post-Doctoral researcher and research faculty at Syracuse University and Mississippi State University. His research interests lie at the intersections of multi-fidelity optimization, evolutionary computing and machine learning, with applications to design and control of autonomous aerospace systems, swarm robotics and resilience of critical infrastructural networks. He has co-authored 48 peer-reviewed journal articles, over 110 full-length (IEEE, AIAA and ASME) conference articles, and 3 book chapters in related topics. His research work has been supported by funding from the NSF, DARPA, ONR, NASA and AFOSR, including the NSF CAREER Award. He is an Associate Fellow of AIAA, a Professional Member of ASME and a Senior Member of IEEE. He is also a member of the AIAA Multidisciplinary Design Optimization (MDO) Technical Committee, where he chairs the Education Sub-Committee, and an elected member of the Design Automation Executive Committee in ASME.