Differentiable algorithms for non-differentiable robotics: contact-rich learning and control
Speaker
Michael Posa
Assistant professor, Mechanical Engineering and Applied Mechanics, University of Pennsylvania.
Title
"Differentiable algorithms for non-differentiable robotics: contact-rich learning and control"
Abstract
As we ask our robotic systems to become more capable, with the ultimate aim of deploying robots into complex and ever-changing scenarios, the vast space of potential tasks drives the need for flexibility and generalization. For all the promise of big-data machine learning, what will happen when robots deploy to our homes and workplaces and inevitably encounter new objects, new tasks, and new environments? A core challenge in generalizable robotics lies in the making and breaking of contact, where non-smooth dynamics clashes with typical assumptions in gradient-based optimization and learning. With the goal of rapid adaptation to novel settings, I'll discuss our progress on real-time multi-contact MPC for dexterous manipulation, where we can realize dynamic motions which dynamically and intelligently make and break contact, with only a simple goal as the given objective.
Control, however, requires a model, and so I will present our recent results on contact-inspired implicit model learning learning, where, by embedding convex optimization, we reshape the loss landscape and enable more accurate training, better generalization, and ultimately data efficiency. Lastly, given time, I'll discuss how model learning and control can synergize in interesting ways: via online adaptation or by synthesizing task-relevant models which identify key aspects of multi-contact dynamics necessary to achieve a goal.
About Speaker
Michael Posa is an Assistant Professor in Mechanical Engineering and Applied Mechanics at the University of Pennsylvania. He leads the Dynamic Autonomy and Intelligent Robotics (DAIR) lab, a group within the Penn GRASP laboratory. His group focuses on developing algorithms to enable robots to operate both dynamically and safely as they interact with their environments. Michael received his Ph.D. in Electrical Engineering and Computer Science from MIT in 2017 and received his B.S. in Mechanical Engineering from Stanford University in 2007. Before his doctoral studies, he worked as an engineer at Vecna Robotics. He received the NSF CAREER Award in 2023 and the RSS Early Career Spotlight in 2023. His work has received awards recognition at TRO, ICRA, RSS, Humanoids, and HSCC.