Article title: 李飞飞's latest CoRL2026 talk transcript: From "closet models" to world models, what kind of AI do robots need?
Body:

Among the speakers at this edition, there is one name that cannot be avoided: 李飞飞. Stanford professor, ImageNet initiator, founder of World Labs. Just before this CoRL, the news that World Labs had been acquired by AMD was fermenting in the industry, and she is 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 participate in a dialogue roundtable. In a sense, this was her "AMD debut."
The person who introduced her was her student and one of the organizers of this CoRL, Yuke (朱玉可). He recounted a detail: at the graduation ceremony, 李飞飞 said to 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 is Atlas, a new model developed by World Labs and the Stanford team: a unified world model for generation and reconstruction.
The following is the on-site transcript brought to you by AI Tech Review.
▎"Atlas: A Unified World Model for Generation and Reconstruction"
It is a great honor to be invited to CoRL, especially since several of my former students are organizers. You did a great job, and I know that because many of my papers and workshop proposals were rejected by you. (Audience laughs) Hats off to CoRL.
Today I want to share a model our lab recently developed, called Atlas: a unified world model for generation and reconstruction.

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 need to solve is: given the state of the world, given a goal, and given the 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 with them. Leiphone

Finally, data is always a challenge. Robotics is a typical 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 what method is used, these three problems remain important in robotics. And they happen to correspond to the problems world models need to 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.
The model I want to focus on today, Atlas, 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 structural signals. It outputs pixel frames; it not only generates frames that are beautiful in space and time, but also synthesizes novel views, thereby achieving high-precision, high-fidelity 3D reconstruction.

Atlas: focused on rendering and simulation
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 that matches the real world. Most methods require a large number of posed images; when pose coverage is insufficient, the reconstructed regions are full of uncertainty.

The other line is generative AI. Diffusion architectures are naturally suited to exploring multi-view generation, bringing us closer and 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 unable to run in real time.

Classic 3D reconstruction struggles to maintain spatial consistency
Changing the viewpoint or running in real time is too computationally expensive
Atlas follows 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 something that connects the two ends: what we call sparse reconstruction. Given very few views—in the extreme case, only 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.

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 is "similar in spirit" but not real: there is a hint of those arcades, but most of it is generated.

With 8 images, the result is much better. Relative to such a large space, 8 images are very sparse, but it basically gives an overview of the entire quad—except that the famous Memorial Church is missing, because it was not in the input images.
Finally, with 33 images, which are still very sparse overall, one of them contains the Memorial Church, so Atlas basically reconstructs a beautiful and complete Stanford Main Quad.
One more point to note: all 33 images were taken at ground level—a person walking around with a camera—but we generated completely different viewpoints, including this aerial view you are seeing now.
What is the technical insight behind Atlas? 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 and generate images with camera poses; take images, estimate poses, and generate more images with new poses based on the input; receive multi-view input and interpolate and extrapolate between views along different camera paths to recover a 3D scene; and also receive depth maps or grayboxes and output RGB pixels, which is very useful in many robotics and gaming scenarios.

Atlas's generation and reconstruction capabilities are very strong
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 of camera-controlled generation across multiple state-of-the-art video models (the Rapidata evaluation agency). Atlas ranked first, clearly ahead of the second place.
Atlas can generate view-consistent videos longer than 60 seconds.

I especially want to compare with our previous-generation model from a year and a half ago: at that time, the generation range and geometric consistency were both very limited. Someone criticized us, saying: "What you are making is not a world model, but closet models." (Audience laughs)
Article title: 李飞飞's latest CoRL2026 talk transcript: From the "cabinet model" to world models, what kind of AI do robots need?
Body:

Now we have moved from the "cabinet" into a much larger world.
This is an interesting generation example: Atlas can take two completely unrelated images, estimate a reasonable camera pose, and then complete the rest of the scene, creating a fantastical world that fuses the two scenes. This is very suitable for creative applications. And this one: given a grayscale box input, Atlas can generate a complete and exquisite large world from it, which is very useful for robotics, design, and creative work.
On the reconstruction side, we compared Atlas with dedicated 3D reconstruction models: on the depth prediction task, tested across 7 datasets, Atlas even surpassed the most recent state-of-the-art dedicated 3D reconstruction and depth prediction models.


Atlas itself does not have to use Gaussian Splatting for representation, but if it is attached downstream, we compared Gaussian splatting reconstruction error with dedicated models on 6 datasets, and Atlas likewise reaches a strong SOTA level. We also adjusted the number of input views, smoothly transitioning from sparse reconstruction to dense reconstruction, and performed well on both datasets.
Let's look at a few examples. This is a famous apartment in the 3D reconstruction community: given only 8 images, Atlas reconstructs this fairly large apartment, and we can sample various viewpoints, with full human control over the camera. This is an example of "single-image upscaling": one input image yields a Gaussian splatting world, a very beautiful fantastical space.


In this example, the model not only generates the scene but also generates depth information, allowing artists to precisely control the camera direction and achieve the desired camera movement.
Here are 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 the dynamics of space-time. Now, with 3 iPhones temporarily set up in our office, Atlas can reconstruct this dynamic space-time scene and present a beautiful bullet time effect.
Another development is that we recently distilled the model to run in real time: 15fps. What you are seeing now is a screen recording of someone "driving" Atlas in real time, generated in real time, and you can see that both fidelity and viewpoint consistency are very good.
(Editor's note: Multi-view autoregressive diffusion plus 15fps real-time generation places self-evident demands on underlying compute throughput. In a conversation right after this talk, 李飞飞 and AMD CEO 苏姿丰 happened to have a deep discussion about the synergy between world models and underlying compute. See today's second article for the transcript.)

Finally, Atlas does have potential applications in robot simulation, and we are testing them: both for training navigation robots and for connecting the real world into simulation and then back to real robot manipulation. These robots are trained only in simulated virtual worlds, yet can be deployed zero-shot to the real world. We are deepening this work with Atlas.

Atlas can reconstruct space for robots and provide simulation environments
You can follow our X account, and you will 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 soon open Atlas to the world, so stay tuned.
Finally, a few announcements. First, the Stanford BEHAVIOR Challenge, now in its second year, with winners announced tomorrow. This year has been very exciting, and we have seen new methods emerge, including generative AI.


In addition, I have many students who worked very hard to get into CoRL, and we have 6 papers. The World Labs team and the Stanford student team put a lot into this research.
Thank you, everyone. Leiphone (official account: Leiphone)
That is the full text of the talk. After this talk titled "A Unified Generative and Reconstructive World Model — Atlas," 李飞飞 also held a fireside chat with AMD CEO 苏姿丰 (Lisa Su) about compute, robots, and "the biggest mistake of this era."
This was the first time the two appeared together after AMD acquired World Labs, but that is another story. See you in the next post.
Leiphone original article, unauthorized reproduction prohibited. See reprint notice for details.
