Maryland Robotics Center Research Symposium 2023

Title:  Move, Manipulate, Multitask, Map: Model-Based Controllers for Versatile Swarms

Robotic swarms are being developed for a range of tasks, including environmental sensing, infrastructure inspection, disaster response, exploration, and agricultural operations. Many of these applications will require swarms to function robustly in uncertain environments. This talk will present model-based approaches to controlling swarms of robots that must rely only on local sensing and signaling, often without global localization, inter-robot communication, or prior data about the domain. Despite these constraints, it is possible to design scalable swarm controllers for a variety of objectives, including spatial coverage, collision-free navigation, cooperative payload transport, task allocation, feature mapping, and scalar field estimation, by applying feedback control and optimization techniques to stochastic and deterministic dynamical models that describe the swarm at different levels of abstraction. The swarm behaviors produced by these controllers will be illustrated in simulations and experiments with small mobile robots. Outstanding challenges and possible future directions for research and applications of swarm control will also be discussed.

Spring Berman is an Associate Professor of Mechanical and Aerospace Engineering and Graduate Faculty in Computer Science and Exploration Systems Design at Arizona State University (ASU). She directs the Autonomous Collective Systems Laboratory and is an Associate Director of the Center for Human, Artificial Intelligence, and Robot Teaming (CHART) within the ASU Global Security Initiative. Dr. Berman received Ph.D. and M.S.E. degrees in Mechanical Engineering and Applied Mechanics from the University of Pennsylvania and the B.S.E. degree in Mechanical and Aerospace Engineering from Princeton University. Prior to joining ASU in 2012, she was a postdoctoral researcher in Computer Science at Harvard University. Dr. Berman’s research focuses on the synthesis of control strategies, including bio-inspired controllers, for robotic swarms and other types of distributed systems, such as soft robot manipulators. Her current research also uses virtual and small-scale physical testbeds to develop autonomous vehicle controllers that mimic human-like driving behaviors. She was a recipient of the ONR Young Investigator Award and the DARPA Young Faculty Award.

Title: A Picture of the Prediction Space of Deep Networks

Deep networks have many more parameters than the number of training data and can therefore overfit---and yet, they predict remarkably accurately in practice. Training such networks is a high-dimensional, large-scale and non-convex optimization problem and should be prohibitively difficult---and yet, it is quite tractable. This talk aims to illuminate these puzzling contradictions.

We will argue that deep networks generalize well because of a characteristic structure in the space of learnable tasks. The input correlation matrix for typical tasks has a “sloppy” eigenspectrum where, in addition to a few large eigenvalues, there is a large number of small eigenvalues that are distributed uniformly over a very large range. As a consequence, the Hessian and the Fisher Information Matrix of a trained network also have a sloppy eigenspectrum. Using these ideas, we will demonstrate an analytical non-vacuous PAC Bayes generalization bound for general deep networks.

We will next develop information-geometric techniques to analyze the trajectories of the predictions of deep networks during training. By examining the underlying high-dimensional probabilistic models, we will reveal that the training process explores an effectively low dimensional manifold. Networks with a wide range of architectures, sizes, trained using different optimization methods, regularization techniques, data augmentation techniques, and weight initializations lie on the same manifold in the prediction space. We will also show that predictions of networks being trained on different tasks (e.g., different subsets of ImageNet) using different representation learning methods (e.g., supervised, meta-, semi supervised and contrastive learning) also lie on a low-dimensional manifold.

Pratik Chaudhari is an Assistant Professor in Electrical and Systems Engineering and Computer and Information Science at the University of Pennsylvania. He is a core member of the GRASP Laboratory. From 2018-19, he was a Senior Applied Scientist at Amazon Web Services and a Postdoctoral Scholar in Computing and Mathematical Sciences at Caltech. Pratik received his PhD (2018) in Computer Science from UCLA, and his Master's (2012) and Engineer's (2014) degrees in Aeronautics and Astronautics from MIT. He was a part of NuTonomy Inc. (now Hyundai-Aptiv Motional) from 2014-16. He is the recipient of the Amazon Machine Learning Research Award (2020), NSF CAREER award (2022) and the Intel Rising Star Faculty Award (2022).