Seminar | June 15 | 4-5 p.m. | 8019 Berkeley Way West
Parsa Mahmoudieh
Electrical Engineering and Computer Sciences (EECS)
Collecting human demonstrations or instrumenting reward functions for learning new tasks is expensive and limits scaling of multi-task policies. In this talk, I will focus on ways to teach and specify new tasks to agents with minimal human supervision. I validate most of my work in the setting of learning perceptual motor tasks as it is a challenging testbed for multi-task learning. My work explores ideas in reusing old demonstrations for learning new tasks with primitive segmentation of data, using curiosity-driven exploration data and self-supervision for learning goal conditioned policies, and leveraging large language vision models for automating language grounded reinforcement learning.
510-982-6207
Jean Nguyen, jeannguyen@eecs.berkeley.edu, 510-642-9413