Li Feifei Team Unveils Atlas World Model for Video, 3D, and Robotics

World Labs, founded by Stanford AI pioneer Fei-Fei Li alongside specialists in computer vision and graphics, has unveiled Atlas, a foundational world model designed to unify video generation, 3D reconstruction, and robotics simulation. Pre-trained from scratch on text, image, video, and 3D datasets, Atlas diverges from conventional video models by treating camera pose as a native input. This mechanism establishes a shared spatial context, allowing the model to comprehend both scene content and precise camera positioning within the environment. The model supports generating up to one minute of 1440p video from a single image and a defined camera trajectory, demonstrating superior trajectory adherence in comparative benchmarks. Simultaneously, Atlas reconstructs 3D environments from sparse 2D inputs, outputting point clouds and Gaussian splatting scenes that integrate directly into existing gaming and visual effects pipelines. Architecturally, Atlas employs a multimodal autoregressive diffusion transformer, merging sequence-based generation with diffusion denoising to process heterogeneous data streams within a unified framework. World Labs positions Atlas as a critical enabler for physical AI and robotics research. By capturing only a few smartphone videos, developers can construct simulated environments where autonomous agents navigate, interact with dynamic objects, and receive synchronized RGB and depth observations. This approach aims to replace specialized scanning rigs with consumer-grade mobile capture, substantially reducing the cost and complexity of real-to-sim data collection. The launch follows a recent $1 billion funding round co-invested by Autodesk, AMD, and Nvidia, highlighting sustained capital inflow into spatial computing infrastructure. Early testing is currently restricted to a select group of industry partners, with independent benchmarks and full model weights yet to be released. Demonstrations validate spatial coherence and visual synthesis but do not yet address complex physics, material dynamics, or the sim-to-real gap. Nevertheless, Atlas represents a decisive industry pivot, moving generative AI beyond frame-by-frame pixel prediction toward true spatial reasoning and comprehensive environmental simulation.