Towards Robots that Learn from Humans

Abstract

How can we make robots that learn new tasks from human teachers? Most research attempts to answer this question by building and testing new imitation learning algorithms. But we believe that the human is equally important: whatever algorithms we develop should stem from how humans teach and interact with robotic systems. This writeup summarizes the insights our group has gained over the last five years by placing humans at the center of the learning problem. We organize our research along three themes, where each theme explores an underlying principle necessary to learn from humans. 1) Learning as control, where we inject structure to align robot learners with how humans teach, 2) learning as representation}, where we enable man and machine to speak the same language, and 3) learning as communication, where we close the learning loop by providing feedback to the human teacher. Viewed together, these interconnected research directions at the intersection of human-robot interaction advance learning from humans in ways that go beyond learning algorithms. Our papers are available at: https://collab.me.vt.edu/