RoboPapers: World Models
How robots build internal representations of the world to predict, plan, act.
A world model is a robot’s representation of its environment. How objects and physical systems behave, and how its own actions might change them. It allows a robot to predict outcomes, plan ahead, and adapt to unfamiliar situations. This is our catalog of foundational papers, important breakthroughs, practical implementations, and credible critiques. Each work is organized by the capability it advances, annotated with its key contribution and limitations, and connected to the research that came before and after it.
Tap to Preview

Visual Geometry Grounded Transformer

  • Ep#8 VGGT: Visual Geometry Grounded Transformer

  • VGGT: Visual Geometry Grounded Transformer

  • GitHub: Code Base

    • 1X World Models

  • Ep#61: 1x World Model

  • Blog: 1X World Model from Video to Action: A New Way Robots Learn

    • DreamZero World Action Models

  • Ep#68: DreamZero: World Action Models are Zero-Shot Policies

  • GitHub Project Page: DreamZero

  • Paper: World Action Models are Zero-shot Policies

    • Multi-stream world action model

  • [Paper] - Flex-π: A Multi-Stream World-Action Model with Compute Flexibility

  • [Interview] - Ep#106: Flex-π: A Multi-Stream World-Action Model with Compute Flexibility