Fall 2023 GRASP SFI: Jim Fan, NVIDIA AI, “Generalist Agents in Open-Ended Worlds”

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

ABSTRACT

Autonomous agents have made great strides in specialist domains like Atari games and Go. However, they typically learn tabula rasa in isolated environments with limited objectives, thus failing to generalize across a wide spectrum of tasks and capabilities. Inspired by how humans continually learn and adapt in the open world, we advocate a trinity of ingredients for building generalist agents: 1) an environment that supports an infinite variety of tasks and goals, 2) a large-scale database of multimodal knowledge, and 3) a flexible and scalable agent architecture. We introduce MineDojo, a new framework built on the popular Minecraft game that features a simulation suite with 1000s of diverse open-ended tasks and an internet-scale knowledge base with YouTube videos, Wiki pages, and Reddit posts. We also propose two new algorithms on top of MineDojo: 1) MineCLIP, a foundation reward function reminiscent of RLHF for embodied agents; and 2) Voyager, an LLM-powered lifelong learning agent that explores and improves itself purely in-context. We look forward to seeing how MineDojo empowers the community to make more progress on the grand challenge of open-ended agent learning.