Embodied AI systems are cyber-physical systems that employ advanced machine learning techniques to perceive, analyze, and interact with their environment. As these systems become more prevalent in fields such as self-driving cars and robotics, ensuring their safety and robustness has become a critical and yet challenging requirement. In this talk, I will discuss key challenges for building safe and robust embodied AI systems, particularly those arising from uncertainties in system inputs, model inaccuracies, and the lack of analyzability of neural network-based components. I will present our recent work on tackling these challenges by developing a holistic set of end-to-end verification, design, and adaptation methods to ensure the safe and robust application of neural networks in the perception, planning, and control of embodied AI systems.
Qi Zhu is a Professor in the ECE Department at Northwestern University. He earned his Ph.D. in EECS from the University of California, Berkeley in 2008 and a B.E. in CS from Tsinghua University in 2003. His research interests include design automation for cyber-physical systems (CPSs) and Internet of Things (IoTs), safe and robust machine learning for embodied AI systems, energy-efficient CPSs, cyber-physical security, and system-on-chip design, with applications in domains such as connected and autonomous vehicles, smart buildings, advanced manufacturing, wearable computing, and robotic systems. He is a recipient of the NSF CAREER award, the IEEE TCCPS Early-Career Award, and the Humboldt Research Fellowship for Experienced Researchers. He received best paper awards at DAC 2006, DAC 2007, ICCPS 2013, ACM TODAES 2016, and DATE 2022. He is the Conference Chair for IEEE TCCPS, and VP of Initiatives for IEEE CEDA. He is an Associate Editor for IEEE TCAD, IEEE TCASAI, and ACM TCPS, and has served as a Guest Editor for the Proceedings of the IEEE, ACM TCPS, IEEE T-ASE, IEEE IoT Journal, Elsevier JSA, and Elsevier Integration, the VLSI journal.