Ben Eysenbach: Contrastive Successor Representations

Feb 12, 2024, 12:30 pm1:30 pm
Joseph Henry room, Jadwin Hall



Event Description

Goal-reaching problems are ubiquitous in both the natural and engineered world. While learning to achieve goals is often considered an aspect of intelligence in biological systems, it is challenging to design practical algorithms for learning such behavior in high-dimensional environments. In this talk, I'll discuss recent work on contrastive successor representations, touching on how these representations might simultaneously facilitate perception, action, and planning. I'll show some results on real-world robots and simulated control tasks.