World Labs just dropped Atlas: An omni world model that simulates space-time and accelerates physical AI
Heat trend
Collecting trend data
The percentage is based on available heat signal, not comment count or independent people.
World Labs has introduced Atlas, an omni-world model designed to advance spatial intelligence and provide AI with a deeper understanding of physics and 3D geometry. This multimodal autoregressive diffusion transformer generates full 3D point clouds and Gaussian splats, rather than just 2D pixels, grounding its outputs in a shared "spatial context." Atlas simulates space-time and is expected to accelerate physical AI development.
The new Atlas model from World Labs is a massive step toward spatial intelligence and giving AI a true understanding of physics and 3D geometry. It’s a multimodal autoregressive diffusion transformer that doesn't just generate 2D pixels, but grounds everything in a shared "spatial context."
The space-time simulation and Real-to-Sim capabilities are the most mind-bending parts of this release:
- The "Holodeck" from a Cell Phone: With footage from just three to five standard cell phones, Atlas builds a complete, navigable space-time simulation. It freezes time and lets you reframe 3D shots from impossible angles without a multi-million-dollar volumetric capture studio.
- Real-to-Sim for Embodied AGI: It doesn't just scan a static room. As a simulated robot moves through a reconstructed space, Atlas actively generates the exact RGB and depth sensor data the robot would see along its specific trajectory.
- Physics and Interaction: It captures how objects actually move and interact in the real world. You can take a casual recording and simulate rigid, articulated, and deformable objects, instantly tweaking the lighting and background to generate infinite training environments.
- Explicit 3D Geometry: Instead of just hallucinating flat video frames, it natively outputs full 3D point clouds and Gaussian splats that can be rendered and explored in real-time.