March 17, 2025, Xingxing Zuo, California Institute of Technology

4:00 p.m.
190 Goodwin Hall

"Towards Spatially Intelligent Robots: Robust 3D Perception for Open-World Autonomy"

Abstract: 3D perception is the cornerstone of spatially intelligent robots, yet achieving consistent robustness remains a significant challenge. Various obstacles arise, including adverse weather conditions, the absence of distinctive structures or textures in environments, highly dynamic robot motion, complex robot-environment interactions, and the limited computational resources and sensing capabilities of deployed robot platforms. In this talk, I will present my research aimed at advancing "Holistic 3D Perception" for mobile robots, with the goal of providing them with robust, comprehensive, real-time awareness of their state and a deep understanding of 3D space in unstructured, real-world environments. I will discuss how my work improves the accuracy, efficiency, and intelligence of 3D perception while addressing the ongoing challenges of robustness. The talk will focus on four key areas: (1) Trustworthy Localization, focusing on achieving unprecedented robustness through on-board computing for reliable operation in challenging conditions; (2) Unlocking Robot Insights, where I address kinematics and spatiotemporal calibration to improve state estimation and sensor fusion; (3) Real-time Incremental Mapping, which integrates sparse tracking with dense reconstruction to enable efficient, real-time 3D mapping; and (4) Understanding Beyond Labels, where I explore open-vocabulary 3D scene representation to push the boundaries of scene understanding. Through these advancements, we strive to move towards a new era of robust 3D perception for autonomous robots, empowering them to perceive, map, and interact with the open world around them.

Bio: Dr. Xingxing Zuo is currently a Postdoctoral Researcher in the Department of Computing and Mathematical Sciences at Caltech. His research interests encompass robotic perception, 3D computer vision, state estimation, and scene understanding. He focuses on advancing highly robust state estimation and 3D space understanding in complex environments, by integrating the intelligence of deep neural networks and probabilistic state estimators. Before joining Caltech in January 2024, Xingxing was a visiting faculty researcher at Google AR (USA). He also held a postdoctoral position at the Technical University of Munich (Germany) from 2021 to 2023. He earned his Ph.D. with honors from Zhejiang University (China) in 2021. During his Ph.D., Xingxing spent one and a half years as an academic guest at ETH Zurich (Switzerland). He was a finalist for the Best Paper Award in Robot Vision at ICRA 2021. Xingxing serves as an associate editor for top-tier robotics publications, including RA-L, ICRA, and IROS, and is a reviewer for numerous top-tier journals and conferences.