Events

Embodied Intelligence Through World Models

Speaker: Danijar Hafner

Location: 60 Fifth Avenue, Room C15

Date: Thursday, March 30, 2023

Deep learning has proven to be a powerful tool for prediction and generation across many domains, including text, images, audio, and video. However, fully automating tasks requires machines to make autonomous decisions, for which traditional algorithms require impractical amounts of data and supervision. Analogous to the success of unsupervised learning in other fields, I argue that the future of decision making will be foremost unsupervised. Towards this vision, I introduce the Dreamer algorithm for learning accurate world models and using them for successful decision making. This algorithm is the first to solve the Minecraft Diamond Challenge from scratch and to teach a physical robot dog to stand up and walk in 1 hour without simulators. Leveraging learned world models, I will then introduce algorithms for autonomous exploration and temporally-abstract decision making that further reduce the amount of supervision. Looking forward, these algorithms pave the way to foundation models for decision making and a broad range of applications.