Anca Dragan's Abstract

Title: An Optimization-Based Theory of Mind for Human-Robot Interaction


Abstract: Generating robot action for interaction with people is not scalable without

learning, but learning from scratch has too high sample complexity. Inductive bias

becomes critical, but what is the right inductive bias when it comes to people? We study

a core assumption: that people are driven by intentions and are approximately rational

in pursuing them. We derive algorithms that can leverage this assumption, ways in which

we can bring the assumption closer to matching real human behavior, as well as ways in

which robots can remain flexible to human behavior that strongly deviates from

rationality.