Vision and Autonomous Systems Seminar
- Newell-Simon Hall
- YEVGEN CHEBOTAR
- Ph.D. Candidate
- University of Southern California
Efficient Learning of Closed-Loop Robotic Skills in the Real World
As robots enter our daily lives they will have to perform a high variety of complex tasks, make sense of their high-dimensional sensory inputs, and cope with highly unstructured, uncertain and dynamic real-world environments. Recent results in robot learning show promise in addressing these problems through learning from experience and adaptation. However, the real-world application of many learning algorithms still remains challenging due to a high sample complexity and difficulty of automating the data collection. In this talk, I will discuss several components that enable efficient learning of sensory-based robotic policies in the real world.
First, I will talk about the ways to decompose the policy learning into efficient training of local components and integrating model-based with model-free learning. Next, I will discuss the concept of self-supervision and external sources of information such as pre-trained reward estimators and reward signals from different domains. Finally, I will talk about moving a large portion of training into simulated environments and using small amounts of real-world experience to adjust distribution of simulated scenarios for a better transferability to the real world. The talk will conclude with a perspective on meta-learning and ways to improve learning performance from previous experience.
Yevgen Chebotar is a Ph.D. candidate at the University of Southern California in the Computational Learning and Motor Control Lab (CLMC) advised by Prof. Stefan Schaal and Prof. Gaurav Sukhatme. Before joining CLMC, Yevgen received his B.Sc. and M.Sc. degrees in Computer Science from the Technical University of Darmstadt. During his Ph.D., Yevgen did a number of internships, including X, Google Brain and Nvidia Robotics. His research focuses on machine learning for robotics, with an emphasis on reinforcement, imitation and transfer learning for efficient acquisition of sensory-based robotic skills.
Sponsored in part by Facebook Reality Labs Pittsburgh