This event will be in-person ONLY in Wu and Chen Auditorium.

ABSTRACT

There are two dominant approaches to designing intelligent robots. One, typified by language behavior models, leverages unstructured deep neural networks and learning from demonstration to generate behavior. These approaches have had several impressive successes but face scaling, trust, and explainability challenges. The second approach seeks to integrate, rather than discard, technologies from existing subfields (like motion planning and SLAM) into a coherent control architecture that retains their favorable properties while accessing the strengths of deep network. The primary challenge here is that there is no unifying theoretical framework for all of robotics: each subfield was designed and studied largely in isolation.

I will propose a unifying framework that models the control stack as layers of increasingly abstract decision processes. Each layer combines perceptual and action abstractions, to generate a more tractable decision process by exploiting structure in the world or the robot. Existing technologies fit naturally into this stack as observation or action abstractions. The result is a natural hierarchy with a few missing technologies. I will discuss my group’s recent results on both filling in these missing technologies, and more generally in learning decision process abstractions from pixel-level data.