ICON Seminar in Robotics: Prof. Soon-Jo Chung (Caltech)
Author: | Tho Le |
---|---|
Event Date: | May 15, 2024 |
Speaker: | Soon-Jo Chung |
Speaker Affiliation: | Caltech |
Time: | 10:30 am-12:00 pm |
Location: | MSEE 112 |
Contact Name: | Tho Le |
Contact Email: | thovle@purdue.edu |
Priority: | No |
College Calendar: | Show |
Time: 10:30-12:00 pm Eastern Time, May 15 (Wednesday), 2024
Location: MSEE 112
Zoom Link: https://purdue-edu.zoom.us/j/98798335169?pwd=ZFNMdmZXendoQ1RCRDczM2dTR1RIdz09
Coffee and snacks will be provided.
Contraction is All You Need in Robot Learning
Abstract:
Deep learning offers necessary representation power for complex dynamical systems, yet it lacks robustness, scalability, and interpretability that are crucial for safety-critical robotic systems. Conventional ML methods require huge amounts of data to train, making real-time re-training of an entire deep neural network infeasible for small robots. Our Neural-Fly (Sci. Robotics, 2022) and subsequent work for fault tolerance overcame those limitations by identifying and rapidly adapting the low-dimensional parameters that must be updated in real time. Its particular adaptive control architecture can also be interpreted from the policy-gradient RL perspective. The key innovation is to systematically guarantee stability and safety in a hierarchically separated architecture using contraction theory. It is no surprise that many well-known convergence results for supervised learning and RL (e.g. policy iteration) rely on exponential stability. I argue that exponential convergence to controllable bounds by contraction, used in conjunction with the comparison lemma, provides superior robustness and stability guarantees for robot learning and control compared to popular methods of asymptotic input-to-state stability. I will present some recent results in the systematic construction of a contraction metric via convex optimization, new safety filtering improving upon control barrier functions, and various real-robot control demonstrations, including LEONARDO, the first bipedal robot capable of walking, flying, slacklining, and skateboarding, and fault tolerant control of urban air mobility vehicles and distributed space robots.
Speaker:
Soon-Jo Chung is the Bren Professor of Control and Dynamical Systems and Senior Research Scientist of the Jet Propulsion Laboratory in the California Institute of Technology. Prof. Chung received the S.M. degree in Aeronautics and Astronautics and the Sc.D. degree in Estimation and Control with a minor in Optics from MIT in 2002 and 2007, respectively. Prof. Chung was an associate professor and an assistant professor at the University of Illinois at Urbana-Champaign. Prof. Chung was a Member of the Guidance & Control Analysis Group in the Jet Propulsion Laboratory as a JPL Summer Faculty Research Fellow and Faculty Affiliate working on distributed small satellites during the summers of 2010-2014. He is the recipient of the UIUC Engineering Dean's Award for Excellence in Research, the Arnold Beckman Faculty Fellowship of the U of Illinois Center for Advanced Study, the AFOSR Young Investigator Program (YIP) award, the NSF CAREER award, a 2020 Honorable Mention for the IEEE Robotics and Automation Letters Best Paper Award, three best conference paper awards (2015 AIAA GNC, 2009 AIAA Infotech, 2008 IEEE EIT), and five best student paper awards. More info: https://aerospacerobotics.caltech.edu/prof-soonjo-chung
Recording:
Organizers: Ziran Wang (ziran@purdue.edu), Nak-seung Patrick Hyun (nhyun@purdue.edu), & Tho Le (thovle@purdue.edu)
2024-05-15 10:30:00 2024-05-15 12:00:00 America/Indiana/Indianapolis ICON Seminar in Robotics: Prof. Soon-Jo Chung (Caltech) Contraction is All You Need in Robot Learning MSEE 112