Talk: Towards building high-performance robotic learning systems

LIVE STREAM: https://uwmadison.zoom.us/j/91461282070?pwd=QWNEZWl0MklWMUtVK2NCNVA3bmtZQT09

Abstract: Robot learning has advanced significantly in recent years, positioning it as an effective tool for achieving scalable, flexible robotic autonomy. However, the large-scale real-world adoption of such learning-based robotic systems remains challenging, for which they must fulfill stringent real-world performance criteria to be viable.

In this talk, I will describe algorithms and principles for building high-performance robotic learning systems. I’ll start by examining a range of high-performance "robot specialist" systems. These systems are tailored to address key deployment factors such as reliability, robustness, and cycle time, which has ultimately paved the way for their industrial adoption. I will then proceed to describe mechanisms to build “robot generalist” foundation models by bootstrapping the aforementioned robot specialists. To conclude, I'll further discuss the connections between these two types of systems and methods for enabling these systems to execute complex, long-horizon tasks suitable for open-world deployment.

Bio: Jianlan Luo is a postdoctoral scholar in the Department of Electrical Engineering and Computer Sciences at UC Berkeley, working with Sergey Levine. Before moving back full-time to academia in 2022, he spent two years as a researcher in the robotics division of Google X. He received his MS/Ph.D. from UC Berkeley in 2020. His research interests lie in the intersection between machine learning, robotics, and controls, with a focus on developing high-performance learning-based robotic systems. His work has led to several industrial adoptions and has been featured in numerous media posts.