Sep 17, 2021 - Michael Posa, University of Pennsylvania

Contact-Rich robotics: Learning, Impact-Invariant Control, and Tactile Feedback

Abstract: Whether operating in a manufacturing plant or assisting within the home, many robotic tasks require safe and controlled interaction with a complex and changing world. However, state-of-the-art approaches to both learning and control are most effective when this interaction either occurs in highly structured settings or at slow speeds unsuitable for real-world deployment. In this talk, I will focus broadly on our most recent efforts to model and control complex, multi-contact motions. Even given a known model, current approaches to control typically only function if the contact mode can be determined or planned a priori. Our recent work has focused on real-time feedback policies using tactile sensing and ADMM-style algorithms to adaptively react to making and breaking contact or stick-slip transitions. For dynamic impacts, like robotic jumping, tactile sensing may not be practical; instead, I will show how impact-invariant strategies can both be robust to uncertainty during collisions while preserving control authority.

In the second half of the talk, I will discuss how such models can be learned from data. While it might be appealing to jump first to standard tools from deep learning, the inductive biases inherent in such methods fundamentally clash with the non-differentiable physics of contact-rich robotics. I will discuss these challenges using both intuitive examples and empirical results. Finally, I will show how carefully reasoning about the role of discontinuity, and integrating implicit, non-smooth structures into the learning framework, can dramatically improve learning performance across an array of metrics. This approach, ContactNets, leverages bilevel optimization to successfully identify the dynamics of a six-sided cube bouncing, sliding, and rolling across a surface from only a handful of sample trajectories.

Bio: 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 computationally tractable algorithms to enable robots to operate both dynamically and safely as they quickly maneuver through and interact with their environments, with applications including legged locomotion and manipulation. Michael received his Ph.D. in Electrical Engineering and Computer Science from MIT in 2017, where, among his other research, he spent time on the MIT DARPA Robotics Challenge team. He received his B.S. in Mechanical Engineering from Stanford University in 2007. Before his doctoral studies, he worked as an engineer at Vecna Robotics in Cambridge, Massachusetts, designing control algorithms for the BEAR humanoid robot. He has received the Best Paper award at Hybrid Systems: Computation and Control (HSCC) and been a finalist for a best paper award at IEEE Humanoids. He has also received Google Faculty Research Award in 2019 and the Young Faculty Researcher Award from the Toyota Research Institute in 2021.

Sep 24, 2021 - Daniel Cohen, Princeton

Oct 8, 2021 - Karthik Narasimhan, Princeton

Nov 5, 2021 - Chuchu Fan, MIT

Nov 19, 2021 - Aaron Ames, Caltech

Dec 3, 2021 - Naomi Leonard, Princeton