The Expert is the Obstacle: Building a General Framework for Learned Robot Motion

Abstract

There are many ways to move a robotic arm through space, but each technique comes with assumptions and trade-offs. Some techniques provide quick, local solutions, while others provide theoretical feasibility guarantees. Furthermore, most state-of-the-art techniques rely on a precomputed scene model to ensure safety. In highly dynamic environments, such as when a robot must operate in a fast-paced industrial setting or around human partners, this assumptions may break down as obstacles move in and out of view. Humans are able to operate effortlessly in these settings, relying on our vast experience to make quick decisions, even under uncertainty. In recent years, we have seen an incredible proliferation of empirical machine learning approaches to long-standing problems, ranging from solving challenging games to producing human-like language. My PhD research has focused on adapting these ideas to robot motion in order to balance the trade-offs of traditional algorithmic approaches to motion generation. In this dissertation, I describe my PhD research on motion generation techniques for robotic manipulation. I will cover my work on trajectory optimization applied to dynamic, multi-agents settings and my research on using large scale imitation learning to create a safe end-to-end policy for manipulator control.