CISE Seminar: October 10, 2019 – Dylan Losey, Stanford University

8 St. Mary’s St., PHO 339
3:00pm-4:00pm

Dylan Losey

Postdoctoral Scholar
Stanford University

Personalizing Robots with Physics and Intelligence

The role of robots in our society is gradually starting to expand. Today, we are interacting with assistive devices to regain motor function and overcome physical disabilities. Tomorrow, we will collaborate with household robots to rearrange furniture, cook dinner, and clean the table. However, without robots that embrace our different behaviors, preferences, and needs, we will never fulfill the promise of autonomy. In this talk, I present a unified formalism for personalized robots that recognize and respond to our individual differences.

My formalism lies at the intersection of physics and intelligence. I view physical control as a means to bring humans and robots closer together, and I leverage cognitive learning to extract human models from interactions. My insight is that personalization will require both physics and intelligence: when robots physically work side-by-side with users, they can intelligently harness interactions to adapt to and influence humans.

I will describe how I have used mechanical compliance and passive control before interactions occur so that people and robots safely work side-by-side. During interactions, I leverage optimal control and Bayesian inference so that intelligent agents can implicitly communicate and exchange roles with one another. The resulting interactions provide a rich stream of information about the human’s individual preferences: after the human stops interacting, I enable robots to learn from human corrections and change their underlying behavior for the rest of the current task. Finally, between interactions the robot should actively adjust its interaction strategy to remove uncertainty and match the human partner.

I have been fortunate to apply my work on personalization to robotic rehabilitation and assistive robotics. I will highlight these applications at the end of the talk. Specifically, I will focus on a control method for assistive robots, where the robot learns a user-friendly embedding between its high-dimensional action space and the user’s low-dimensional input space. Results from user studies on robotic arms demonstrate the effectiveness of this personalized approach.

Dylan Losey is a postdoctoral scholar at Stanford University within the Artificial Intelligence Lab, where he studies human-robot interaction. Dylan received his Ph.D. in Mechanical Engineering from Rice University in December 2018, his M.S. in Mechanical Engineering from Rice University in May 2016, and his B.E. in Mechanical Engineering from Vanderbilt University in May 2014. Between May and August 2017, he was also a visiting scholar at the University of California, Berkeley, where he worked in the Berkeley Artificial Intelligence Research Lab. Dylan researches human-robot interaction at the intersection of physics and intelligence, and leverages machine learning and control theory. He was awarded the 2018-19 Rice University Outstanding Thesis Award in Mechanical Engineering, the 2017 IEEE/ASME Transactions on Mechatronics Best Paper Award, and was a National Science Foundation Graduate Research Fellow.

Faculty Host: Wenchao Li
Student Host: Mahroo Bahreinian