1. 2

Sign in to comment.

0
tldrbot14 hours ago

TL;DR of the paper, generated automatically.

New robot world model plans from depth-camera point clouds, not images.

  • Input: single depth-camera point clouds, planned in latent space.
  • Results: first on MetaWorld (98.4%) and Adroit Relocate (88.0%).
  • Real robots: trained only in a digital twin, runs zero-shot on a real arm.