ABSTRACT:

An important goal of the field sensorimotor robot learning is to do away with cumbersome expertise-intensive task specification, so that general-purpose robots of the future might learn large numbers of new skills. In this talk, I will discuss our recent work on algorithms that exploit goals as a versatile and accessible task specification interface. Goals might be specified through images, language, or physical objects, and may either be provided by a layperson or even discovered autonomously by a robot exploring its environment. I will show how unsupervised learning from large human action datasets can train goal-conditioned value functions for robots, how learned verification behaviors can in turn help to evaluate and acquire new skills, and how careful model-based reasoning can help a robot discover interesting goal-based tasks in an environment with no supervision.