Fei-Fei Li's CoRL 2026 talk transcript: World Labs releases Atlas, a unified world model for generation and reconstruction
In her CoRL 2026 talk, Fei-Fei Li introduced Atlas, developed by World Labs and the Stanford team: a multi-view autoregressive diffusion Transformer trained from scratch that uses next-view prediction to unify generation and reconstruction within a single model. Atlas can generate pixel frames with camera poses from text, image, video, depth map, and grayscale box inputs, supports switching from dense reconstruction to sparse reconstruction to pure generation, can generate view-consistent videos longer than 60 seconds, and has been distilled to run in real time at 15fps. She said Atlas is being used for robot simulation testing and will soon be opened to the world.
Why read this
AI score 67
At CoRL, Fei-Fei Li explained how Atlas uses a single autoregressive model to unify generation and reconstruction, and shared progress on real-time operation and robot simulation.
Topics
Body · AI translation
Two years later, World Labs' work is already completely different.
English text is being prepared.
English text is being prepared.
Among the speakers at this edition was one unavoidable name: Fei-Fei Li. Stanford professor, originator of ImageNet, founder of World Labs. Just before this CoRL, news that World Labs had been acquired by AMD was fermenting within the industry, and she was about to become AMD's chief scientist. At this CoRL, she not only gave a comprehensive introduction to World Labs' recent work and her industry views, but also invited AMD's Su to take part in a dialogue roundtable—in a sense, this was her "AMD debut."
Introducing her at the opening was her student and one of the organizers of this CoRL, Yuke (Zhu Yuke). He recounted a detail: at the graduation ceremony, Fei-Fei Li told him, "When you were in my lab, you were my student, but from now on, you are my friend."
The formal protagonist of this talk was Atlas, a new model developed by the World Labs and Stanford teams: a unified generative and reconstruction world model.
▎Speaker: Fei-Fei Li
It is a great honor to be invited to CoRL, especially since several of my former students are organizers. You did a great job—I know, because many of my papers and workshop proposals were rejected by you. (Laughter) Hats off to CoRL.
Today I want to share a model our lab recently developed, called Atlas: a unified generative and reconstruction world model.
Since we are at CoRL, let me start with the robotics motivation.
The field of robotics has changed a lot recently, but no matter what method we use, the fundamental problem robots must solve is: given the world state, given a goal, given a robot, what should be done next at every moment?
Broadly speaking, robots need to solve several problems: first, understand and represent the state of the world—whether implicitly or explicitly is up for discussion, but "understanding" itself matters; second, predict actions with consequences and plan accordingly. Leiphone
Finally, data is always a challenge. Robotics is a classic long-tail problem, and successful deployment is extremely limited, so we always hope to supplement real data by simulating as many different environments as possible, for training and evaluation.
No matter the method, these three problems remain important in robotics. And they happen to correspond to the problems a world model must solve: understanding the environment corresponds to rendering; predicting the consequences of actions corresponds to planning; simulating world states corresponds to simulation.
Rendering, planning, and simulation: the three major dimensions of world models
I don't want to dogmatically say they are completely independent; in fact, in many respects they can be unified and connected to one another, and these are the most cutting-edge and exciting directions in this field.
Today I will focus specifically on Atlas, which mainly addresses rendering and simulation, especially simulating the geometry of the world and 3D/4D dynamics.
In short, Atlas is an autoregressive foundation model trained from scratch, a single model that receives multimodal inputs: text, images, video, depth signals, graybox, or 3D structure signals. It outputs pixel frames; it not only generates frames that are beautiful in space and time, but can also synthesize novel views, thereby achieving high-precision, high-fidelity 3D reconstruction.
Atlas: focused on rendering and simulation
Let me first review a bit of history.
Computer vision has two very important problems that researchers have spent generations trying to understand.
One is the 3D reconstruction problem, from the classic "Building Rome in a Day" to recent NeRF-style neural rendering work. The fundamental problem is: given images from different camera poses, build a 3D world consistent with the real world. Most methods require a large number of images with poses; when pose coverage is insufficient, the reconstructed regions are full of uncertainty.
The problem with classic 3D reconstruction
The other line is generative AI. Diffusion architectures are naturally suited to exploring multi-view generation, bringing us ever closer to "generation + reconstruction." There have been many excellent works attempting to jointly denoise multiple views to maintain spatial consistency, but overall the results are limited: spatial consistency is hard to maintain, and each denoising pass corresponds only to a fixed target view, making it computationally expensive and impossible to run in real time.
Classic 3D reconstruction struggles to maintain spatial consistency
Changing views or running in real time is too computationally expensive
Atlas continues this line of thought, attempting to unify generation and reconstruction under one roof with a single model architecture and a single training recipe.
Here is how it works. First, one extreme: given multiple images, Atlas performs reconstruction; when there are enough images, it performs dense reconstruction and recovers the scene well. The other extreme: zero pixel conditioning, with only a text prompt, Atlas performs pure generation and generates multiple frames of view-consistent images.
Atlas can also do things connecting the two ends: what we call sparse reconstruction—given only a very small number of views, in the extreme case just one image, Atlas can estimate camera poses, then generate as many views as possible, reconstructing and generating a complete world. Within a single model, it switches freely from dense reconstruction to sparse reconstruction to pure generation.
Let's look at an example.
At least some of you will recognize this: it is the famous Main Quad on the Stanford campus. Given two limited images, the model "hallucinates" a Main Quad. Those familiar with Stanford know this is not the real Main Quad—it "resembles" it but is not: there is a hint of those arcades, but most of it is generated.
Given 8 images, the result is much better. Relative to such a large space, 8 images are very sparse, but it basically gives the overall appearance of the entire plaza—except that the famous Memorial Church is missing, because it is not in the input images.
Finally, with 33 images, it is still very sparse overall, but one of them contains the Memorial Church, so Atlas basically reconstructed a beautiful, complete Stanford Main Quad.
One more point to note: these 33 images were all taken at ground level—just one person walking around with a camera—but we generated completely different viewpoints, including this aerial viewpoint you are seeing now.
What is Atlas's technical insight? It is next-token prediction, which everyone has heard of.
Our formulation is: Viewpoint is all you need. Predicting the next viewpoint under strong camera conditioning—that is our objective function. Architecturally, it is a multi-view autoregressive diffusion Transformer.
So Atlas can: take text input and generate images with camera poses; take image input, estimate poses, and generate more images with new poses based on the input; receive multi-view input and interpolate and extrapolate between viewpoints along different camera paths to recover a 3D scene; and also receive depth maps or grayscale boxes and output RGB pixels, which is very useful in many robotics and gaming scenarios.
Atlas's generation and reconstruction capabilities are very strong
Let's look at some quantitative results.
A few days ago, a third-party organization we had not known about in advance used about 500,000 human judgments to conduct a blind test (by the Rapidata evaluation organization) of camera-controlled generation across multiple state-of-the-art video models. Atlas ranked first, clearly ahead of second place.
Atlas can generate videos over 60 seconds long with consistent viewpoints.
I especially want to compare with our previous-generation model from a year and a half ago: at that time, generation range and geometric consistency were both very limited, and someone criticized us: "What you're making is not a world model, it's a closet model (secret, not public—like a model shut in a closet)." (Laughter)
Now we have walked out of the "closet" into a much bigger world.
Here is an interesting generation example: Atlas can take two completely unrelated images, estimate reasonable camera poses, and then complete the rest of the scene, creating a fantastical world that fuses two scenes. This is very suitable for creative applications. And this one: given grayscale box input, Atlas can generate a complete, exquisite large world from it, which is very useful for robotics, design, and creative work.
On reconstruction, we compared Atlas with specialized 3D reconstruction models: on depth prediction tasks tested across 7 datasets, Atlas even surpassed the latest state-of-the-art specialized 3D reconstruction and depth prediction models.
Atlas itself does not have to use Gaussian Splatting as its representation, but if it is connected downstream, we compared Gaussian Splatting reconstruction error with specialized models on 6 datasets, and Atlas likewise achieves a strong SOTA level. We also adjusted the number of input views, smoothly transitioning from sparse reconstruction to dense reconstruction, performing well on both datasets.
Let's look at a few examples. This is a famous apartment set in the 3D reconstruction community: given only 8 images, Atlas reconstructs this rather large apartment, and we can sample various viewpoints, with full human control over the camera. This is an example of "lifting a single image into dimensions": one input image yields a Gaussian Splatting world, a very beautiful fantastical space.
In this example, the model generates not only the scene but also depth information, allowing artists to precisely control camera direction and achieve the desired camera movement.
Let's look at two more fun ones. People familiar with film know "Bullet Time": that classic visual effect in The Matrix, which back then required an array of hundreds of cameras to capture spatiotemporal dynamics. Now, with 3 iPhones temporarily set up in our office, Atlas can reconstruct this spatiotemporal dynamic scene and present a beautiful bullet-time effect.
Another development is that we recently distilled the model to run in real time: 15 fps. What you are seeing now is a screen recording of someone "driving" Atlas in real time, generating in real time, and you can see that fidelity and viewpoint consistency are both very good.
(Editor's note: Multi-view autoregressive diffusion plus 15 fps real-time generation places self-evident demands on underlying compute throughput. In the conversation right after this talk, Fei-Fei Li and AMD CEO Lisa Su happened to discuss in depth the coordination between world models and underlying compute—see today's second article for the full transcript.)
Finally, Atlas does have potential applications in robot simulation, and we are testing it: both for training navigation robots and for connecting the real world into simulation and then back to real robot operation. These robots are trained only in simulated virtual worlds, yet can be deployed zero-shot to the real world. We are deepening this work by combining it with Atlas.
Atlas can reconstruct space for robots and provide simulation environments
You can follow our X account and soon see the premiere of this work; I promise it is a great video. Creators can use Atlas to realize their creative dreams, with precise camera control, object interaction, and real-time navigation. We will open Atlas to the world very soon, so stay tuned.
Finally, a few advertisements. First, the Stanford BEHAVIOR Challenge, now in its second year, with winners announced tomorrow. This year is very exciting, and we have seen new methods emerge, including generative AI.
In addition, I have many students who worked hard to get into CoRL; we have 6 papers, and the World Labs team and the Stanford student team put a lot of effort into this research.
Thank you all. Leiphone
▎Postscript:
The above is the full text of the talk. After this talk titled "A Unified Generative and Reconstructive World Model—Atlas," Fei-Fei Li also had a fireside chat with AMD CEO Lisa Su about compute, robots, and "the biggest mistake of this era."
This was the first time the two appeared on stage together after AMD's acquisition of World Labs, but that is another story—see our next post.