In the first half of 2026, money and talent in Chinese AI moved in one direction. World models.
Disclosed funding in the domestic world model category reached roughly 30 billion yuan, about 4.2 billion dollars, in six months. Fold in the combined embodied intelligence and world model category and the figure swells past 90 billion yuan, around 12.6 billion dollars. Against the same period last year, that is growth of more than five times.
Which raises an awkward question. Of everyone rushing in, how many have actually worked out what the thing is?
Practitioners are looking for direction, investors are looking for people, headhunters are mapping teams, and corporate strategy departments are trying to work out who survives. Everyone is asking the same thing. Can anyone explain world models clearly?
Zhang Lei can. He is a chair scientist at IDEA Research and founder of Vision Future, an IEEE Fellow with more than 77,000 Google Scholar citations and an H-index of 107. His DINO series topped the COCO leaderboard for five months, and Grounding DINO and the later DINO-X beat Google and Meta to become standard citations for Fei-Fei Li’s group, NVIDIA and Galbot. He spun Vision Future out of IDEA in 2025, with an angel round of close to 100 million yuan landing within a month. His academic mentor is Academician Zhang Bo, one of the founders of Chinese AI, and his industry mentor is Harry Shum, former executive vice president of Microsoft and former head of Microsoft Research Asia.
In a year when everyone is arguing about the definition, Zhang is one of very few people giving a clear, operational answer that refuses to compromise with any of the fashionable camps.
Why embodied AI is still short of the line
Robots have to solve two things, Zhang says. Success rate, meaning whether a task completes reliably, and generalisation, meaning whether the same task still completes when the environment or the body changes.
He lists four concrete gaps.
First, weak generalisation inside a single scenario. Plenty of teams now reach 70 or even 80 per cent success in one setting, which sounds like a breakthrough. In real use, 80 per cent means one failure in every five attempts, which still needs a large human backstop and cannot be deployed independently.
Second, insufficient breadth across scenarios. The right order is to push a single capability to its ceiling first, then extend sideways into more environments and harder tasks. That gap is the largest obstacle between laboratory demonstration and real deployment.
Third, missing depth in the brain. Deep learning today is largely learning the behaviour of intelligence rather than its mechanism. An embodied brain needs genuine causal inference about the physical world, not statistical correlation but the logic of because A therefore B, plus common sense and adaptation. That work has barely started.
Fourth, nothing close to edge deployment. Most robots still lean on large external compute. Until the need for a human backstop disappears, the loop of application, data and iteration cannot form.
Asked whether hardware or the brain is the real bottleneck, Zhang picks the brain without hesitation. Chinese hardware iteration is a genuine source of pride, he says. Companies like Unitree and AgiBot have refined motors and transmission through simulation-based reinforcement learning and cut cost and improved reliability substantially. But what those bodies show in marathons and performances is roughly the baseline capability of an animal. Instinctive reaction, jumping, dodging and physical interaction with the environment are still out of reach.
The difficulty with the brain, he says, is that we do not understand the mechanism. Every breakthrough in deep learning has looked at intelligence from the surface and then implemented it with algorithms. AI is learning human behaviour, not human intelligence. As Zhang Bo often says, a large language model can talk like a person, but the way it understands language is completely different from the way a person does.
Why the concept came back after twenty years
World models are not new. Researchers were modelling the environment an agent interacts with in the early 1990s to improve reinforcement learning algorithms, but deep learning had not arrived and the idea was almost impossible to validate.
Validation came around 2018 and 2019, when papers actually titled World Models appeared and were applied in games, which are the best environment for testing reinforcement learning. Once AI could beat human players, researchers began asking whether an agent could learn a world model on its own through interaction, packing the evolution of the environment into the model. That idea was fully validated between 2018 and 2020.
The link to embodied AI runs through language models. The field first built vision language action models, borrowing the imitation learning paradigm. A language model predicts the next token, and an embodied VLA looks at the current frame and predicts the next action.
That method hit a wall quickly, and Zhang argues data is only part of it. The bigger part is that reinforcement learning is underused. His analogy is neat. Imitation learning taught language models to speak. Reinforcement learning taught them to speak correctly. Robots need the same second step for success rate.
But reinforcement learning on physical robots faces two lethal obstacles. Data collection efficiency is extremely low, far below language, because the robot must physically execute each action and wait for feedback. And failure cost is unacceptable. Fail at washing dishes and you break the dish. Learn to avoid car accidents and you may need a lot of real accidents.
The virtual environment a world model provides is the way out. If a model can predict what the environment becomes after a given action, the robot can rehearse thousands of failures in its head rather than breaking ten thousand actual bowls.
The definition: action conditioned, or it does not count
Asked what a model needs to deserve the name, Zhang gives one condition. Action dependence. The model must be able to answer what the environment becomes if a specific action is executed.
An agent interacts with its environment through actions. The action changes the environment, the change produces a new state, and that state tells the agent whether it moved closer to the goal. Every state transition depends on the specific action taken in the previous step.
People often describe a world model as replacing next token prediction with next state prediction. Zhang calls that definition sloppy because it drops the key condition. The correct phrasing is action conditioned next state prediction. Only with the action as a precondition does state prediction mean anything for reinforcement learning.
By that standard, pure video generation models such as early Sora do not qualify. They do not model action at all. They generate a self-consistent video sequence and predict how pixels change, without modelling the interaction between robot and environment. Such a model can teach a robot some rules about how the world changes, but it does very little for interaction.
The newer world action model category also fails his test. It uses pixel-level supervision from predicted future frames to improve action prediction and beats VLA on some tasks, but its modelling runs effect before cause, generating the future frame and then inferring the action backwards, which cannot be used for reinforcement learning.
Latent space versus pixels, and the missing third option
The loudest fight in the field is between Yann LeCun’s latent space camp and the pixel camp that Sora represents.
Zhang starts by pointing at what they share. The apparent opposition partly reflects LeCun stating his position with unusual clarity, because a researcher who wants a position remembered has to state it sharply. But the pixel route also relies on latent space. Stable Diffusion and Sora both compress images or video into a lower-dimensional representation before processing. So representation space is the more fundamental question. The real disagreement is not whether to use latent space, but what to keep and what to discard once inside it.
The latent route wants to strip out lighting and texture, pixel detail not driven by physical law, which helps the model learn the deeper rules behind state change. Its risk is representation collapse. If ten thousand different frames map to an identical representation, the model has lost the ability to distinguish states. The pixel route pursues perfect visual fidelity, which is fundamentally about pleasing the human eye, because people are extremely sensitive to visual detail.
Asked whether human mental simulation runs in latent space, Zhang agrees immediately. If the human brain has a world model, imagining how an action changes the environment happens in latent space, not by rendering the whole scene pixel by pixel. Judging a world model by human perception of image quality is therefore inadequate. If the pixel route wants to build world models, the target should not be photorealism but physical plausibility across the generated sequence.
His own difference from LeCun is one word. Structure. They share the same first principle, that human counterfactual reasoning is action conditioned and runs in abstract representation rather than replaying pixels. LeCun validated the axiom that prediction should happen in representation space through I-JEPA, V-JEPA and LeJEPA. But physical space in embodied settings is extremely complex, and Zhang argues that using the method at scale in real environments requires introducing necessary structure into the latent space.
That view comes from capability he has already proven. His core team led DINO-X, which solved open world object perception. So now the team makes objects the basic unit of prediction and planning. Physical law acts on objects, he says, so making the latent representation understand objects lets the model learn physical law more efficiently.
Asked why that improvement can outrun LeCun’s own AMI Labs, which has raised well over a billion dollars for the same direction, Zhang is direct. Because we built DINO-X, a general open-set object perception model, and we made understanding objects the best in the world. Our understanding of why vision models produce physical plausibility maps closely onto the general object understanding line we have already run. LeCun’s route is the base framework. We introduce object structure into it and close the loop in real environments, and that closing ability is what we have built over the last several years.
How you test whether a model understands anything
Understanding is hard to verify directly, Zhang admits. Nobody can explain how a language model understands language, only judge it from behaviour. If it outputs text you cannot distinguish from a person, you say it understands.
World models sit in the same trap. Whether a model has genuinely captured physical law, or object structure, cannot be stated cleanly. His team analyses intermediate steps through visualisation, and if the result matches human intuition, that suggests object structure has been captured.
There is a sharper test. Counterfactual testing. Let the model try different actions and see whether the result is physically plausible. If a slightly different trajectory still produces the right outcome, the model has learned a rule. If it fails, what it learned was probably statistical correlation.
That takes discipline. As Richard Sutton says, researchers cannot resist helping the model by pouring their own experience in. Language model researchers once wanted to feed in syntactic structure, tokenisation and phrase parsing, and they still lost to simply predicting the next token. Zhang says he strongly believes in the power of simple paradigms iterated forward. Excessive human design works in the short term and blocks iteration in the long term.
On the objection that sand, water, smoke and cloth have no clear boundaries and might break an object-centric framework, he points to masks. Segmenting an object region in pixels is the simplest entry, and vision has become mature at semantic object understanding and mask prediction, including non-rigid material. Starting from masks is a pragmatic entry point, and our accumulation in general object perception is deeper than any team outside.
Action alignment, and why it matters
A world model has two inputs, the frame and the action, which makes understanding action critical.
Action alignment is the basic requirement for handling multiple bodies. Aligning a human hand with a robotic arm is the harder case. A robot arm’s joint positions are precisely known in three-dimensional coordinates, while human hand data is usually two-dimensional video. Someone films themselves cooking and the recording contains no three-dimensional coordinates at all, so mixing the two directly is very inefficient.
The team extracts three-dimensional key points, converting two-dimensional human hand video into a manipulation sequence in three-dimensional space aligned to the same physical frame as arm data. That is the egocentric data problem the industry now talks about, and the team’s earlier EgoTwin release addressed exactly that, currently in collaboration with Baidu Cloud’s data collection group.
Higher level intent, the question of why the hand moved that way, gets left to the world model or the VLA, but only if data quality is high enough. Raw material is critical, Zhang says. The industry pours enormous headcount and cost into collection, and in many cases the data is used inefficiently. The core lesson from Grounding DINO and DINO-X is to put people who understand algorithms deep into the data work, otherwise what you collect may be unusable at training time.
Editor’s note: This is an adapted translation of the original Leiphone interview. It has been trimmed and restructured for readability for an international business audience.