We present Rapid Exploration Controllers for Outcome-driven Navigation, or RECON, a simple framework that combines latent goal models with a non-parametric memory for exploration in open-world environments with minimal human supervision.

An extended version of this talk was presented at the 2022 Conference on Robot Learning (CoRL) in London, UK.

"RECON: Rapid Exploration for Open-World Navigation with Latent Goal Models"
Dhruv Shah, Gregory Kahn, Nicholas Rhinehart, Sergey Levine
Berkeley AI Research (UC Berkeley) and Carnegie Mellon University

Project Page: [ Ссылка ]
arXiv pre-print: [ Ссылка ]

свернуть