9:00 AM – 12:30 PM, February 24, CSL B02
As robots become increasingly integrated into our daily lives, there is a need to study robot learning and control algorithms to guarantee stability, performance and robustness: ideas that have been studied in control theory for decades. The design of theoretically performant control techniques often informs algorithms in robotics. This session will aim to explore topics at this intersection of theory and practice, highlighting advances in robotics driven by theoretically rigorous algorithms.
Keynote Speaker – Prof. Dimitra Panagou, University of Michigan
Time: 9:00 – 10:00 AM
“Reliable and Resilient Multi-Robot Systems”
Abstract: Resilience of a networked system against failures and attacks is a fundamental property that allows the system to “keep running” even at reduced functionality. Research on resilience and related concepts such as fault tolerance involve distributed learning, planning and control approaches and has been vivid across multiple communities including robotics and control. Multi-robot systems in particular pose significant challenges given that attacks or failures can occur either at the “cyber” domain, e.g., the information shared via communication among robots or acquired via onboard sensing, or at the “physical” domain, e.g., at the vital components of each robot (sensors, actuators) or to the entire system as a whole. Despite tremendous progress, there are still open problems, including but not limited to how we can obtain less conservative responses to possible failures and attacks beyond those that correspond to worst-case robustness, and how we can ensure the safe, secure and efficient operation of the robots in challenging environments, e.g., subject to internal and external constraints. In this talk, I will present an overview and highlights of our recent work on resilient multi-robot systems with a focus on safety, security and efficiency.
Biography: Dimitra Panagou received the Diploma and PhD degrees in Mechanical Engineering from the National Technical University of Athens, Greece, in 2006 and 2012, respectively. In September 2014 she joined the Department of Aerospace Engineering, University of Michigan as an Assistant Professor. Since July 2022 she is an Associate Professor with the newly established Department of Robotics, with a courtesy appointment with the Department of Aerospace Engineering, University of Michigan. Prior to joining the University of Michigan, she was a postdoctoral research associate with the Coordinated Science Laboratory, University of Illinois, Urbana-Champaign (2012-2014), a visiting research scholar with the GRASP Lab, University of Pennsylvania (June 2013, Fall 2010) and a visiting research scholar with the University of Delaware, Mechanical Engineering Department (Spring 2009). Her research program spans the areas of nonlinear systems and control; multi-agent systems, autonomy and control; and aerospace robotics. She is particularly interested in the development of provably-correct methods for the safe and secure (resilient) operation of autonomous systems in complex missions, with applications in robot/sensor networks and multi-vehicle systems (ground, marine, aerial, space) under uncertainty. She is a recipient of the NASA Early Career Faculty Award, the AFOSR Young Investigator Award, the NSF CAREER Award, the George J. Huebner, Jr. Research Excellence Award, and a Senior Member of the IEEE and the AIAA.
Industry Speaker – Dr. Jian Yao, XPENG
Time: 10:00 – 10:30 AM
“XPENG Iron: Beyond the most Human-like CatWalk Humanoid”
Abstract: Humanoid Robot has attracted lots of attention from both academia and industry, thanks to the breakthrough in the large foundation models, e.g. VLA, World Model, etc. On one hand, awesome demos often exposed to social media make people believe humanoid robot is very close to commercialization. On the other hand, people who are actually working in this area think there are still some fundamental problems we need to solve before the humanoid robot can widely enter our life. As an important milestone, XPENG aims at humanoid robot production at the end of the year 2026, serving the use cases like tour guides, shopping guides and guided patrol. To achieve this goal, there are a lot more challenges we need to overcome than just a demo show. In this talk, instead of sharing the solution, I will talk more about problems we need to solve for the production, including but not limited to robot ‘AEB’, robot vibration on navigation system, computation power constraint for multimodal large language models, thermal issue, data scarcity and the robot agent challenge.
Biography: Jian Yao is currently the Director of AI at XPENG Robotics, where he focuses on robotics foundation models and their applications in navigation and manipulation. Prior to joining XPENG, Jian worked at Apple and NVIDIA. He received his Ph.D. in Computer Science from the University of Toronto in 2016. His earlier research centered on applying cutting-edge machine learning techniques to address challenges in vision, language, and speech.
Invited Student Speaker – Qiyang Li, UC Berkeley
Time: 10:30 – 11:15 AM
“Reinforcement learning with Action Chunking”
Abstract: Recent progress in robotic manipulation policy learning has been largely driven by (1) the increasing availability of large-scale prior datasets and (2) the success of action chunking, where the policy predicts a short sequence of future actions rather than a single one. However, most action chunking policies are trained via supervised imitation learning, because efficient online self-improvement with reinforcement learning (RL) remains challenging—limiting their real-world adaptivity. In this talk, I will discuss my recent work on leveraging prior data to optimize action-chunking policies with RL, combining empirical results with theoretical insights. I will first present Q-chunking, a recipe that leverages Q-learning and flow-matching generative models to enable sample-efficient online fine-tuning of action chunking policies. I will then share theoretical insights that inform us when and how we should use action chunking with RL, and conclude with an outlook on leveraging large-scale prior data for more efficient online RL beyond action chunking policies.
Biography: Qiyang (Colin) Li is a PhD student at UC Berkeley advised by Prof. Sergey Levine. His research interests include reinforcement learning and robot learning, with a focus on leveraging offline prior experience for online exploration. Before that, he was an undergraduate student at the University of Toronto advised by Prof. Roger Grosse.
Student Presentations
Time: 11:20 AM – 12:30 PM
Gia Quoc Bao Tran: “Stability of Slow-Fast Nonlinear Systems”
Anant Joshi: “Error Analysis of Sampling Algorithms for Approximating Stochastic Optimal Control”
Sandeep Banik: “Shared Autonomy: From Robots to AI agents”
Yan Miao: “Training Once, Flying Freely: The FalconGym Ecosystem for Zero-Shot Sim-to-Real Aerial Autonomy”
Andre Schreiber: “Do You Know the Way? Human-in-the-Loop Understanding for Fast Traversability Estimation in Mobile Robotics”
Shivansh Patel: “Robotic Manipulation by Imitating Generated Videos Without Physical Demonstrations”
Runpei Dong: “HERO: Learning Humanoid End-Effector Control for Open-Vocabulary Visual Loco-Manipulation”