This will be a hybrid event with in-person attendance in Levine 307 and virtual attendance on Zoom.

ABSTRACT

The lack of large robotics datasets is arguably the most important obstacle in front of robot learning. While large pretrained models and algorithms like reinforcement learning from human feedback led to breakthroughs in other domains like language and vision, robotics has not experienced such a significant influence due to the excessive cost of collecting large datasets. In this talk, I will discuss techniques that enable us to train robots from very little human feedback, as little as one demonstration or one language instruction, or their natural eye gaze. I will dive into reinforcement learning from human feedback, and propose an alternative type of human feedback based on language corrections to improve data-efficiency. I will finalize my talk by presenting how existing large pretrained vision-language models can be used to generate direct supervision for robot learning.