Ben Mildenhall

Why World Models Could Change Robotics, 3D, and Creativity

Fei-Fei Li, Justin Johnson, Ben Mildenhall, Martin Casado

World Labs has unveiled Atlas, a unified spatial intelligence model that uniquely combines generative video with sparse 3D reconstruction to create high-fidelity simulations from as few as three input views. This architectural breakthrough enables "AI-complete" new view prediction and robotic real-to-sim transfer while reducing data capture requirements by up to 100 times compared to traditional dense methods. By treating 3D poses as native inputs, the system delivers consistent, dynamically stable environments that eliminate the unpredictability of standard generative video for applications ranging from industrial design to autonomous robot training.