Picture this. You pull out your phone and snap a few photos of your living room. A dozen seconds later, a virtual robot is standing inside a digital twin of that same room, working out how to move around the coffee table, push open the balcony door, and carry a package from the entrance to the desk.
This is not a scene from a science fiction film. It is what Fei-Fei Li’s team just delivered.
On September 2, 2026, World Labs, the spatial intelligence company Fei-Fei Li founded, released a new world model called Atlas. The company was not shy about the label it slapped on it: “the first multimodal world model in the world.” The claim is that a single model can handle video generation, 3D reconstruction, and robot simulation all at once. In the past, those three jobs belonged to three completely separate technical tracks, each with its own dedicated teams and toolchains.
Atlas wants to pull all of them under one roof.
The spatial brain behind a single photo
To understand why Atlas caused such a stir, you first have to look at an awkward truth in today’s video generation field.
Over the past two years, video generation models have been in a frenzy of competition. Sora, Kling, Seedance took turns on stage, and every few weeks a new model claimed sharper images, longer clips, higher resolution. But one problem never really got solved: how the camera moves is basically guesswork.
You type “camera slowly rotates left, slight upward tilt” into the prompt box. What does the model do with that? It is really just guessing among a pile of pixels what the next frame probably looks like. As for how many degrees “left” means, where the center of “rotation” sits, how far up “upward tilt” goes, it has no idea. Get lucky and the shot comes out smooth. Get unlucky and the image warps and falls apart.
Atlas did something fundamentally different.
It turned the camera’s position and orientation in three-dimensional space into a native input for the model. The technical term is Spatial Context. Put plainly, Atlas not only knows “what is in this photo,” it also knows “from which point in space, facing which direction, this photo was taken.”
What does that buy you? It means when you tell it “move the camera along this path,” what you hand over is no longer a fuzzy text description but a precise string of three-dimensional coordinates. Atlas generates frames one by one along those coordinates, keeping the spatial geometry consistent from start to finish.
According to 36Kr’s report, you only need to feed Atlas one to six ordinary photos, set the camera path, and it can output video up to a minute long at 1440p. The stranger part is its ability to fill in space. Give it a photo showing only the front of a robot, swing the camera around to the back, and it can reconstruct the rear view and surrounding environment that never appeared in the input.
That filling-in is not conjured from nothing. During training, Atlas took in a large volume of real-world images and 3D data, building up a set of prior knowledge about “what the world usually looks like.” When it sees a corner of a garden, it infers from that knowledge that beyond the fence there is probably a lawn and distant hills, then fills in the missing parts. The more viewpoints you feed it, the less the model has to imagine. MIT Technology Review’s Chinese edition put it well: two or three photos usually yield a fairly faithful reconstruction, while Atlas can use more than a hundred photos as spatial context.
To be fair, once the camera moves into an area the original photos never covered, the model is still essentially guessing. The wider the viewpoint gap, the more likely local geometry drifts and textures flicker. What it generates is more like a visually plausible three-dimensional world, not necessarily an exact digital replica of that world.
From “the image looks good” to “the space is real”
If Atlas only shot steadier video, it would count as one more iteration in the video generation race. What really made the industry sit up is the 3D reconstruction ability it showed at the same time.
Traditional 3D reconstruction has a steep barrier. Whether Gaussian splatting or point cloud scanning, you need professional gear circling the target dozens of times, shooting hundreds or thousands of photos. Atlas compressed that process to a jaw-dropping degree. In the Stanford Main Quad case the company showed, it used only a few dozen ground-level tourist-angle photos to restore the full detail of the lawn, the arcade, and the memorial church facade, then lifted the camera straight up into an aerial flyover from a bird’s-eye view.
On public benchmarks, Atlas also turned in strong numbers. Quantum Bit reports that when measuring sparse-view reconstruction error on test sets like DTU, ETH3D, and ScanNet, Atlas scored 25.3, beating Pi3X’s 28.7 and Depth Anything 3’s 39.3. On this metric, a lower score means more accurate reconstruction.
More importantly, what Atlas outputs is not just a video that “looks like 3D.” It can predict depth for every frame, assemble the results into a point cloud, and go further to generate a 3D Gaussian splatting scene that drops straight into a game engine, a VFX tool, or architectural design software. Generation and reconstruction are no longer two separate pipelines but two outlets of the same model.
World Labs also pulled out a “bullet time” demo as its killer showcase. Without Hollywood’s ring of hundreds of synchronized cameras, the researchers used only three to five ordinary phones and action cameras, shooting a dynamic scene from a few directions at once. After Atlas realigned those viewpoints and built a spatial model, a user could rewatch the same instant, after the event, from camera angles that do not exist in reality. Recreating the famous Matrix shot with five phones was unthinkable two years ago.
The real goal is the robots
Now, if you think Fei-Fei Li went to all this trouble to fight Hollywood for special-effects work, you are underestimating her.
Atlas’s deepest ambition points at the biggest bottleneck of this generation of AI, embodied intelligence, getting robots not just to “chat” but to actually walk into the physical world and do work.
Anyone who has trained robots knows a brutal reality: training data is expensive. Letting robots try and fail repeatedly in a real factory is slow, costly, and carries safety risk. Train in a simulated environment instead? Then you first have to spend a lot of labor building realistic 3D scenes, setting materials and lighting, placing objects. There is an estimate floating around the field that collecting robot training data at sufficient scale through traditional means could end up costing an astronomical amount.
Atlas’s answer is blunt. Shoot a video of a real environment with your phone, feed it to the model, and it spits out a high-precision 3D digital twin space. The robot climbs in and starts training.
The clever part is that Atlas did not just build a static set. Wherever the robot walks, it renders in real time the RGB image and depth data the onboard camera should see. When the arm bumps a box, pulls a cabinet door, or squeezes a sponge, it simulates the corresponding feedback. Record one real scene and you can change the lighting, rearrange the obstacles, and spin off a large number of different training scenarios.
World Labs calls this path Real-to-Sim, from reality to simulation. After seeing Atlas, Nvidia’s robotics lead Jim Fan called it a key step for Real-to-Sim in robotics.
Pull the timeline back and Atlas does not look sudden. This June, Fei-Fei Li personally wrote an essay dividing world models into three types: the renderer that handles the image, the simulator that handles the physical environment, and the planner that handles decisions, and she stated plainly that “the most critical one is the simulator.” In July, World Labs acquired the robot simulation company SceniX one after another and unveiled a Real-to-Sim-to-Real system. By the time Atlas arrived in September, this strategic roadmap could not be clearer.
The technical foundation: a hybrid of four paradigms
Atlas can do three jobs in one model thanks to a fairly elegant underlying architecture.
The company gave it a somewhat clunky name: multimodal autoregressive diffusion Transformer. Unpack it and you find several of the most mature technical routes in AI today blended together.
“Multimodal” means text, images, video, camera pose, and 3D depth maps are all native inputs, no extra adapters or bridging modules required. “Autoregressive” is the classic large language model move, predicting the next token from what came before, except Atlas predicts not the next word but the next spatial state. “Diffusion” borrows the denoising mechanism from image generation models, ensuring image quality and geometric precision for high-dimensional visual signals. The underlying “Transformer” architecture lets it reuse the KV cache, distributed inference, and other engineering optimizations that the LLM field has been polishing for years.
This hybrid architecture brings a very practical benefit: different tasks are really just different arrangements of the input and output sequence. Want to generate video? Put the images and camera path up front, output video frames after. Want to do 3D reconstruction? Put multi-angle photos up front, output depth maps and point clouds after. Want to build a training ground for a robot? Put the environment video and robot pose up front, output what it should see after. Same model, same weights, and switching tasks only means rearranging the sequence.
In blind tests of camera control, Atlas performed strongly. After third-party annotators evaluated it, Atlas’s win rate against Seedance 2.5 reached 94%, 81% against Gemini Omni Flash, and 93% against FLUX 3. Still, MIT Technology Review pointed out a detail worth watching: Atlas can directly accept precise camera trajectory parameters, while its rivals can only accept text-described camera instructions. That both reflects Atlas’s design advantage and means this was not entirely a comparison under equal conditions.
A cool head: how far from “simulating reality”
Atlas is exciting, but equating it directly with “a physical simulator that can predict reality” is premature.
The robot cases published so far mainly verify spatial reconstruction, visual generation, and simple object interaction. On harder physical layers like collision, friction, material properties, and complex dynamics, World Labs has not given systematic test data. The more pressing question: how do robot policies trained in Atlas-generated simulation actually perform back in the real world? That gap, known as the sim-to-real gap, is the core problem the whole field acknowledges, and Atlas has not answered it yet.
On top of that, Atlas has not disclosed its parameter count, training data volume, or compute used, and has not open-sourced its weights. It has just entered early testing with a small number of partners, and the published evaluation results come mainly from World Labs itself, without independent third-party verification.
None of that stops us from seeing an increasingly clear direction.
Fei-Fei Li once compared today’s large language models to “wordsmiths in the dark,” able to talk about the world but never having actually lived in it. A model can explain why a cup falls off a table, but for a robot to actually reach out and grab the cup, it needs to know how far the cup is from the edge, from which direction the hand reaches over, and how the whole scene changes once it touches the cup.
Language models handed machines the text knowledge of the internet. What world models do next is hand them three-dimensional space, physical law, and temporal change as well. Atlas is not the finish line, but it may be the leg of this relay that has run the farthest so far.
Look back at Fei-Fei Li’s two decades, from ImageNet teaching machines to recognize objects in pixels, to founding World Labs so machines understand the three-dimensional space behind those pixels, and she has always been chasing the same question: after a machine sees an image, does it really understand the world?
Atlas says maybe it is starting to.
Related reading
- The Real Lock-In Was Never the Model: AI Memory Can Now Move With You
- Best Serverless Databases for Next.js Projects in 2026: Real-World Testing
- Amazon Is Shutting Down Mechanical Turk: The Machine That Hid Humans Inside Finally Doesnt Need Them
- Clay Alternatives for B2B Data Enrichment: A Real 2026 Comparison
- The Real Barriers Blocking Enterprise AI Agent Adoption in 2026
Stay updated with our latest AI insights
