DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning

摘要

We introduce DINO-WM, a world model built on top of pre-trained DINO visual features that enables zero-shot planning without task-specific training. By leveraging the rich semantic representations learned by DINO, our world model can predict future visual states and plan actions for novel tasks directly from visual observations. DINO-WM achieves strong zero-shot performance on manipulation and navigation tasks across diverse environments.