This is a hybrid event with in-person attendance in Levine 307 and virtual attendance via Zoom. This week’s presenter will be in-person as well.

ABSTRACT

While it is tempting to view robotics as a nail that can be solved with the deep learning hammer, we have seen that deep-learning based perception and action pipelines for robots are notoriously brittle and data hungry. In this talk, I advocate for a more measured approach for designing data-driven controllers by focusing learning on task-relevant portions of the MDP. Through this philosophy, I show that we can acquire capable learning systems that can transfer between morphologically distinct robots, intelligently probe the environment for imperceptible reward signals, and perform deep exploration with no priors.