Date: September 30, 2026
Speaker: Ruoshi Liu, University of Maryland
Title: Small World Models
Host: Kuan Fang
Abstract: World models is one of the hottest topics today in robot learning. But why are they not the de facto standard tool for contact-rich manipulation? In this talk, I will talk about the major limitations of the existing world models that make them not feasible for contact-rich and dexterous manipulations and propose a new paradigm forward, which I termed “Small World Models”. Rather than relying on internet-scale pretraining and human data, small world models focus their capacity on real-world physical interaction collected natively by robots, showcasing surprising long-horizon, reactive, and physically consistent results, as well as strong scalability. I will present several recent works, along with preliminary results, that point toward this direction.
Bio: Ruoshi Liu is an assistant professor of computer science at University of Maryland, College Park, as well as an Amazon Scholar at Amazon FAR. Previously, he was a research scientist at Meta FAIR and completed his B.S., M.S., and Ph.D. at Columbia University. His research aims to develop intelligent systems that can perceive and interact with the physical world, with a special focus in robot manipulation.
