World Labs, founded by Fei-Fei Li and others, announced its next-generation world model "Atlas" on September 1, 2026. Instead of asking a video AI in text to "pan left," it directly passes a camera's position and orientation in three-dimensional space, generating both the images visible from that trajectory and the 3D shape with the same model. From 1 to 6 reference images, it can produce up to a minute of video at 1440p, and it also outputs point clouds and 3D scenes. The premise Atlas shifts is a move away from treating video as a time-ordered sequence of pixels, toward conditioning the model on where each pixel was seen from.
However, there's a gap between clean-looking video and physics simulation that a robot can trust. The demos and company evaluations World Labs published show Atlas is strong at camera control and 3D reconstruction from sparse images. Whether objects move correctly in response to actions, and whether real robot policies can be safely evaluated, aren't published as standalone Atlas figures. The value of this announcement becomes clearer once you separate what's been demonstrated from what remains unverified.
Spatial Context: Placing Images in 3D Coordinates
What Atlas receives, in addition to text and images, are camera poses and depth maps. Video is represented as a sequence of images, with each image and its depth tied to a shooting position and orientation. This combines multiple inputs into a single "spatial context," which then conditions what viewpoint's image or 3D information the model generates next.
The foundation is a multimodal autoregressive diffusion Transformer. Like a language model, it processes sequence elements one at a time, and for image and depth generation it uses a rectified-flow-type diffusion process that progressively removes noise. Rather than switching to a different model for each task, the design changes how the sequence is arranged—what is output after the input—to handle camera-controlled generation, 3D reconstruction, and spatiotemporal reconstruction.
Explicit camera pose brings generation closer to a "shooting instruction." Text-based descriptions of pans, crane moves, and zooms are open to interpretation, but camera coordinates and orientation let you fix the target trajectory numerically. From a single photo, Atlas generates a specified viewpoint, inferring the back of objects or adjacent spaces not visible in the input. Add a second or third real photo, and the inferred regions get replaced by observations. World Labs says even more than 100 images can be fed into the same spatial context.
This mechanism calls for opposite evaluations depending on the use case. For film and games, the ability to plausibly fill in an unseen hallway or garden shortens production time. For surveying, architectural documentation, and robot training, that same completion can amount to hallucinated geometry. Being able to build a wide world from a few photos and being able to accurately measure a real space are not the same thing.
Inside the Evaluation: Unifying Video and 3D Reconstruction
After generating images or video from new viewpoints, Atlas can output that world as depth, point clouds, and 3D Gaussian Splatting (3DGS). 3DGS represents space using numerous semi-transparent Gaussian distributions, making it easy to render at high resolution and frame rate even on-device. Since it's the same representation used in World Labs' existing product Marble, generated results can be readily passed into web display or production pipelines.
In the company's camera-control evaluation, third-party raters compared pairs of videos to judge which better followed a specified trajectory. Atlas was preferred 75% of the time against MiniMax H3, 81% against Gemini Omni Flash, 86% against Happy Horse 1.1, 93% against FLUX 3, and 94% against Seedance 2.5. The gap reportedly widened as trajectories grew more complex.
But this test's input conditions aren't symmetric. Atlas was given camera trajectories in a native format, while competing models were given text describing cinematic terms like pans and crane shots. World Labs itself acknowledges that competitors might improve with better-crafted prompts. So the 75–94% preference rates should be read less as a ranking of overall image quality and more as a combined measure of a design that directly accepts camera geometry plus the model's ability to follow trajectories.
For 3D reconstruction, Atlas predicts the 3D point corresponding to each pixel from images and camera pose, and was compared against specialized models across seven public datasets. Average AbsRel error was 25.3 for Atlas, 28.7 for Pi3X, 34.7 for π³, 36.4 for VGGT-Ω 1B, 39.3 for Depth Anything 3, and 47.7 for MapAnything, in units of 10⁻³, where lower is better. World Labs says it re-ran the baselines using a common procedure. However, this is a company-organized evaluation, not a result accompanied by a technical paper or independent replication.
For Robotics, Separating the World You See from the World That Moves
In June 2026, World Labs split world models into three functions. The renderer produces the "observations"—such as images—that sensors receive. The simulator produces the "state," including geometry, physics, and dynamics. The planner outputs "actions" toward a goal. The company itself distinguishes between video optimized for appearance and a state in which objects, when force is applied, move according to physical laws.
Atlas crosses the boundary between renderer and simulator, but it hasn't been presented as a planner that outputs actions. While functioning as a renderer that produces pixels from new viewpoints, it also produces explicit state in the form of depth, point clouds, and 3DGS. In a robot navigation example, the company captured 24 frames per environment from smartphone video and generated RGB images and depth as seen by a virtual robot's onboard camera. In another demo, events captured by 3 to 5 smartphones or action cameras were reconstructed from a different angle with time frozen.
On the other hand, the inputs currently listed for Atlas are text, images, camera pose, and depth—robot actions are not included. Regarding manipulation tasks, World Labs' own wording is that Atlas "assists" in building simulations. It explains that scenes containing rigid bodies, articulated objects, and deformable objects can be varied, but it doesn't present friction, contact, deformation error, or real-world policy transfer rates as evaluations of Atlas itself. There's also no basis for calling Atlas a planner that handles action generation.
This distinction becomes even clearer when looking at the R2S2R engine, derived from SceniX, that World Labs published in July. A separate system linking Real-to-Sim (converting real environments into simulation environments) and real-world transfer reportedly matched physical interactions and transferred policies trained purely in simulation to real robots. In policy evaluation, each checkpoint was tested 2,000 times in simulation and 100 times on real hardware, examining rankings and correspondence between failure regions. This is more rigorous verification than the Atlas announcement provides, but those results can't be transferred to Atlas's standalone performance.
The product roadmap is starting to converge. In December 2024, World Labs released a research demo that builds an explorable 3D world from a single image, and in November 2025 it made Marble generally available. In January 2026 it opened the World API, integrated SceniX in July, and in September presented Atlas as a foundation model. Atlas is set to power a future version of Marble, playing a role in aligning the entry points for creative 3D generation and robot simulation under the same spatial representation.
Atlas's Strength Is Narrower—and Sharper—Than "World's First"
If we go only by the classification of "multimodal" or "omni," world models existed before Atlas. In May 2026, NVIDIA announced the open Cosmos 3, which handles text, images, video, ambient audio, and actions. Odyssey's Starchild-1, released the same month, is a research preview that generates audio and video in real time while responding to streaming inputs of text, speech, and actions. Google DeepMind's Genie 3 generates an environment users can move through for several minutes at 720p, 24fps.
What should be compared isn't the number of modalities but which part of the world each model makes computable. Among the four models examined, only Atlas simultaneously published native input of camera pose and depth alongside explicit output of point clouds and 3DGS.
| Model | Input/control emphasized in the official announcement | Explicit output/role | Public availability as of September 2, 2026 |
|---|---|---|---|
| World Labs Atlas | Text, images, camera pose, depth | Images, video, depth, point clouds, 3DGS | Early access for selected partners |
| NVIDIA Cosmos 3 | Text, images, video, ambient audio, actions | World generation, physical reasoning, action trajectories | Model and code released |
| Google DeepMind Genie 3 | Text and user movement controls | Interactive video world at 720p, 24fps | Research prototype |
| Odyssey Starchild-1 | Continuous input of text, speech, actions | Real-time generation of synchronized audio and video | Research preview |
What sets Atlas apart among the four models examined is the combination of taking camera pose and depth as native input while returning generated pixels and explicit 3D shape to the same coordinate system. Cosmos 3 extends a different axis with actions and openness, Genie 3 with real-time interactivity, and Starchild-1 with synchronized audio. Rather than broadly calling Atlas's strength "the first multimodal model," defining it more narrowly as "a model that outputs both shootable video and computable 3D from a single foundation" more accurately captures the technical distinction.
What Early Access Conceals: The Metrics Still Needed Before Real-World Use
Atlas has entered early access for selected partners. As of September 2, 2026, Atlas isn't listed in a public API, and no paper, model card, pricing, or general-availability date can be confirmed. The announcement page also doesn't reveal parameter count, training data breakdown, inference speed, generation cost, or partners. Despite a development organization backed by a newly raised $1 billion in February 2026, the product conditions users would need in order to reproduce results still aren't in place.
In production work, what will be tested is whether shape and texture hold up when the same scene is generated repeatedly along different camera paths, and how far editing can go once 3DGS is passed into Unreal Engine or Three.js. For architecture and real-environment reconstruction, a mechanism is needed that separates the creativity used to fill in unseen regions from measurement error. In robotics, the deciding factor won't be how RGB and depth look, but whether contact, friction, deformation, and post-action states match a real robot.
Whether Atlas truly becomes a shared foundation for production and robot development depends less on when it joins a public API and more on whether third parties can reproduce the accuracy of its camera and geometry, and whether that accuracy can be linked into a single closed loop with SceniX's physics and policy evaluation. If those conditions are met, the process of manually building separate virtual sets for filming and separate worlds for robot training could turn into computation branching from the same spatial context.