Talk Abstract
In this talk, I will take you on a journey from the fundamental stages of imitation learning to the realms of personalized robotic companions. I will illustrate how we challenged the status quo, proving that robots could effectively navigate and manipulate their environment using only RGB data and training solely in simulation. This breakthrough led us to recognize that imitation is not just powerful but also generalizable to the real world.
I will then showcase our efforts in harmonizing navigation and manipulation into a seamless, end-to-end solution. Our approach enables robots to efficiently and adaptively perform complex household tasks that necessitate simultaneous navigation and manipulation, such as opening doors and cleaning tables.
Concluding the talk, I'll introduce 'Promptable Behaviors', an innovative yet straightforward approach to robot personalization. Here, I will demonstrate how we've utilized multi-objective RL to enable robots to adapt their behaviors in real time, aligning with diverse human preferences and objectives without the need for retraining. Join me to discover how our findings could shape the future of robotics, potentially making them not just tools, but adaptable and personalized companions in our everyday lives.
Biography
Kiana Ehsani is a Research Scientist at PRIOR (Perceptual Reasoning and Interaction Research) at the Allen Institute for AI (AI2). Before joining AI2, she completed her Ph.D. at the Paul G. Allen School of Computer Science at the University of Washington, under the guidance of Professor Ali Farhadi. Her research primarily focuses on the intersection of computer vision, machine learning, and embodied AI, with a special emphasis on sim2real transfer and imitation learning for domestic robots.
Website: https://sites.google.com/view/ehsanik-personal-website
Location
Sennott Square Building, Room 5317
Date
Friday, February 16 at 2:00 p.m. to 3:15 p.m.
Faculty Host
Dr. Adriana Kovashka