2026 Winter Robotics Colloquium: Florian Shkurti (University of Toronto)

Описание: Title: Evaluating policies without breaking your robots
Speaker: Florian Shkurti (University of Toronto)
Date: Friday, February 20, 2026

Abstract: As our field becomes increasingly better at training multi-task manipulation policies, concerns regarding how to comprehensively and safely evaluate these policies are becoming more prevalent. In this talk I will present my group's recent projects on off-policy evaluation of diffusion policies based on compositional stitching; policy evaluation via action-conditional video models; and policy evaluation via generation of transferable adversarial scenarios. A key theme will be how to make the most out of data we already have, and leverage imperfect learned models to enable policy evaluation for manipulation. I will also outline promising research directions going forward.

Bio: Florian Shkurti is an Assistant Professor at the Department of Computer Science at the University of Toronto, where he directs the Robot Vision and Learning Lab. He's spending his sabbatical year as a Visiting Research Scientist at Ai2 in Dieter Fox's robotics group. He is a faculty member at the University of Toronto Robotics Institute, the Vector Institute, and the Acceleration Consortium for lab automation. His research broadly spans robotics, machine learning, and computer vision. He has received the Alexander Graham Bell Doctoral Award, the AAAI Robotics Fellowship, the Amazon Research Award in Robotics, the Connaught New Researcher Award, the TRI Young Faculty Researcher award, and the AI2050 Fellowship from Schmidt Sciences.

This video is in the process of being closed captioned.