Andrey Kolobov

About

I am a Principal Researcher and research manager at MSR Redmond, working on sequential decision-making with a particular emphasis on robotic manipulation. My current research revolves around the following overarching questions: (1) How can robots use the prior knowledge encoded in multimodal models as well as physical interaction-specific inductive biases for data-efficient learning to act during (lifelong) post-training? (2) What type of knowledge should we pre-train into these models to enable this kind of efficient decision learning? Despite focusing on robotics, I believe that the very same questions under different inductive biases are crucial to unlocking the potential of non-robotic and even non-embodied decision-making agents, including those based on language-only large models.

In the past, I worked on topics ranging from decision-making and atmosphere modelling for autonomous sailplane UAVs to scheduling algorithms for Bing’s next-generation Web crawler to understanding the solution properties of MDPs with irrecoverable failure states.

I received a CS Ph.D. from the University of Washington – Seattle and a double B.A. in CS and applied mathematics from UC Berkeley.