Speaker:Professor Siddhartha Srinivasa The Robotics Institute Carnegie Mellon University
Seminar Venue:Executive Classroom, COM2-04-02
For years, the focus of robot motion planning has been to produce functional motion: industrial robots move to weld parts, vacuuming robots move to suck dust, and personal robots move to clean up a dirty table. We have been exploring the thesis that although functional motion is ideal when robots perform tasks in isolation, it is insufficient for collaboration, where a human and a robot are manipulating in a tightly-coupled shared workspace. Our goal is to make this collaboration fluent and seamless. To this end, we have been developing algorithms where the notion of an observer watching the motion is woven into the fabric of the motion planner. This perspective has allowed us to formalize qualitative notions such as predictability and legibility in psychology in terms of Bayesian inference and inverse optimal control, and to develop generative models for such motion using functional gradient optimization. I will also describe some of our user studies on applying these algorithms to human-robot handovers, assistive teleoperation, and shared workspace collaboration, and ongoing work on deception, ambiguity, and emotive motion.