In 2026, "world model" is one of the least agreed-upon terms in the AI world.
A video that can be generated continuously is called a world model; a digital environment that changes in real time with keyboard and mouse operations is also called a world model; systems that predict future states in latent space, as well as policies that directly output robot actions, likewise use this name.
These models all process information about the world, yet their capability boundaries are far apart. The images in generated videos can be realistic enough, yet violate real physical laws; a robot can complete a grasp, but it may also only be familiar with the objects, camera positions, and motion trajectories in the laboratory. Whether a system has learned visual correlations, physical structure, or the relationship between actions and outcomes is hard to answer from a demo alone.
Conceptual confusion increases the difficulty of technical judgment. Image quality, geometric precision, prediction capability, and task success rates are mixed together for comparison; different teams appear to be competing in the same track, but may actually be answering different questions.
Therefore, world model research is beginning to return to a basic question: what capabilities must a model have in order to move from generating content to understanding and changing the world?
Li Feifei and World Labs divide world models by function into renderers, simulators, and planners, which output pixels, world states, and actions, respectively.
Link: https://www.worldlabs.ai/blog/taxonomy-of-world-models
LeCun, by contrast, advocates learning predictable world structures in an abstract latent space, allowing the model to retain information useful for understanding, reasoning, and planning.
Link: https://ai.meta.com/research/vjepa/
Zhu Jun of Tsinghua University offers a third angle. As early as last December, when Motus was released, he publicly laid out the overall vision of a "general world model"; in March of this year, he positioned it as the foundation connecting the digital world and the physical world. A few months later, Zhu Jun and his team further refined this framework in the paper "General World Models from First-Principles," defining core capabilities from first principles and providing a five-level evolutionary roadmap.
Paper link: https://www.shengshu.com/assets/gwm-principles-and-roadmap.pdf
Technical report: https://www.shengshu.com/zh/general-world-model/
The paper summarizes the core capabilities of a general world model as understanding, imagination, and action. A model needs to form a judgment about the current world from historical observations, simulate the possible futures of different choices, take actions, and then use new observations to correct itself.
Along this framework, video generation, real-time interaction, latent-space prediction, and robot control can be regarded as different stages on the same capability path.
How far world models are from general intelligence is also transformed into more specific questions: Can it maintain a coherent world? Can it predict the consequences of interventions? Can it learn from real feedback? And can it gradually form goals and organize more complex actions?
The First Principles of World Models
Many people have had the experience of learning to ride a bicycle.
At first, the body always wobbles. Seeing the handlebars tilt to one side and feeling the body's center of gravity begin to sink, the person immediately adjusts direction, and then continues to correct based on new feedback. With more practice, the brain gradually learns to anticipate what result a certain action will bring.
This is precisely the most basic meaning of a world model.
In 1943, Kenneth Craik proposed that humans form "small-scale models" of reality in their minds, using them to simulate the possible outcomes of different actions. POMDP in reinforcement learning further describes this process: the agent receives partial observations, estimates the world state, takes actions, and then updates its judgment based on new observations.
Zhu Jun's team summarizes this logic as the three core capabilities of a general world model, which connect with one another and form a continuously updating closed loop:
Understanding: integrating information such as vision, language, sound, touch, or robot sensors into an internal judgment about the current world;
Imagination: simulating multiple possible futures from the current state, especially "what would happen if I took action A instead of action B";
Action: turning predictions into interventions in the digital or physical environment, and letting the new observations after action in turn test and correct the model.
The three core capabilities of a general world model form a closed loop: understanding is responsible for inferring the current state, imagination is responsible for predicting possible futures, and action changes the world and provides new evidence for the next round of understanding.
This is also why video generation is not the entirety of a world model.
Video models can answer what the next
Laihua
scene might look like, but to move toward a general world model, it also needs to answer questions like "If I change this condition, move this object, or apply this action, how would the world be different?" Only predictions with action conditions like these touch upon causality, counterfactuals, and executable decision-making.
Understanding, imagining, acting—how do they form a closed loop?
To describe how world models evolve, Zhu Jun's team further proposed a five-level roadmap, namely from L1 generating the world, L2 interacting with the world, and L3 acting in the world, gradually moving toward L4 autonomous world agents and L5 world organizers.
It is more intuitive to understand this with a glass cup pushed to the edge of a table.
An L1 model can generate a continuous sequence of the cup sliding off, hitting, and shattering; an L2 model can continue evolving this world after the user changes the viewpoint or the pushing direction; by L3, the model needs to judge whether the cup will fall, predict whether reaching out will be in time, and output a grasping action that a robot can actually execute. Only after acting will it know whether its judgments about the cup's weight, friction, and grasping timing were correct.
Moving further up, L4 no longer just waits for instructions from people; it must be able to discover risks, actively observe, supplement information, and revise strategies after failures. L5 further faces multi-agent collaboration, such as who grabs the cup, who avoids obstacles, who allocates tools, and how to rearrange tasks after the environment changes.
The significance of this roadmap is that it turns "world model" from a noun into a set of questions that can be tested level by level.
L1 looks at whether generated trajectories are coherent, L2 looks at whether the world can respond continuously, L3 looks at whether the model can change the physical environment, L4 looks at whether the system can actively explore and continuously learn, and L5 looks at whether it can organize multi-agent collaboration in open environments.
The formal modeling goals and acquired capabilities of the five-level roadmap for general world models (GWM).
According to the judgment given in the paper, existing systems have already reached the first three levels, while L4 and L5 remain open problems. The main gaps include causal and physical grounding, persistent memory, online learning, efficient deployment, and safety control. Evaluation must also compare predictions with real outcomes and examine the model's transferability to unfamiliar tasks, environments, and embodiments.
From video to embodiment, two validations of the same hypothesis
The difficulty of world models is not only in algorithms.
Internet video is massive in scale and records a wide range of objects, people, scenes, and motions. Models can learn spatial structure, object permanence, human behavior, and how events usually unfold from it, but video rarely simultaneously records the actions, forces, and intentions that cause changes. Robot trajectories can connect observations, actions, and outcomes, but collection costs are higher, and they are easily tied to specific hardware and tasks.
The path proposed by Zhu Jun's team is to place these two types of data into a data pyramid that moves from observation to action.
The data pyramid of general world models (GWM) and its relationship to the five-level capability roadmap.
At the bottom is internet-scale video, used to acquire world knowledge and dynamic priors. Above it, in order, are domain video, first-person human video, human demonstrations with action records, and at the top, real robot interaction data. The higher up one goes, the scarcer and more costly the data, but the clearer the correspondence among actions, tasks, and robot embodiments.
This explains why Shengshu Technology advances both video generation and embodied intelligence: video provides the breadth of world knowledge, while robot data completes action grounding.
On the digital world side,
Vidu
series continues to explore generation and interaction with the visual world. Video is not the endpoint of world models, but it is an important starting point, allowing models to learn how objects, scenes, motions, and events evolve amid large-scale spatiotemporal changes.
Vidu
S1 further advances one-shot generation to real-time response, allowing user input to continuously change subsequent content and testing whether the model can maintain state continuity during interaction.
On the physical world side, Motus and Motubrain connect environment understanding, state prediction, and robot actions into the same pipeline. The model needs to see the environment clearly, predict the consequences of intervention, then translate judgments into executable actions and subject them to validation by real outcomes.
Existing representative systems that realize L1 to L3.
The two product lines thus form two validations of the same world model hypothesis. The digital environment tests whether the model can understand and imagine the world, while the physical environment tests whether this knowledge can support action and be revised through feedback.
To enable the two types of verification to share underlying capabilities, images, video, language, and robot actions need to compute collaboratively around the same world state. Traditional modular systems pass information level by level, which easily loses geometric, temporal, and uncertainty information at the interfaces. Fully sharing parameters across all modalities may also cause training conflicts, letting massive video data suppress scarce robot signals.
The MoT (Mixture-of-Transformers) architecture proposed by Zhu Jun's team attempts to strike a balance between the two. Different modalities retain their own expert parameters, then exchange context through a shared attention mechanism. Vision tells the model what is happening in the environment, language provides goals and constraints, and action brings intervention into the environment; the three types of information jointly update the model's judgment about the world.
MoT provides the foundation for unified computation, but physical laws, cross-embodiment transfer, real-time inference, and safety control still need to be solved through data, training, and systems engineering. The goal it points to is very clear: a general world model needs a world state that can be jointly read and written by different modalities and continuously updated.
Motus2:
Let robots, before acting,
first try it out in their own world
If Motus and Motubrain advanced Shengshu's world model exploration to L3, enabling robots to generate actions based on environmental understanding and future prediction, then the recently proposed Motus2 further adds outcome evaluation and policy feedback, attempting to give the action loop self-improvement capability.
Motus2 does not make action generation, future simulation, and outcome evaluation into three mutually independent systems. Instead, it uses a single video-action model with shared parameters to expose three control interfaces:
Policy (Policy / World-Action Model): proposes executable candidate actions;
Simulator (Simulator / Action-Conditioned World Model): predicts the possible consequences of these actions in the visual world;
Evaluator (Evaluator / Value Model): judges which outcome is closer to the task goal, and uses this judgment for selection and improvement.
You can think of it as a true case of "first the mind, then the hands."
The robot no longer simply gives an action directly after seeing an object. Instead, it first proposes several candidate plans, rehearses in the model the images and task progress each would bring, and then chooses the better step. After execution, success, failure, and less-than-ideal results are no longer just data waste, but become evidence for the next round of dynamics modeling and value learning.
Action, prediction, evaluation, feedback, and re-action thus enter the same decision and learning loop, which is also the closed-loop self-evolution emphasized by Motus2.
On the data side, Motus2 adopts a hierarchical training path, expanding from monocular first-person video to synchronized binocular first-person data, and then completing embodiment adaptation through robot data. Its training uses about 130,000 hours of human first-person recordings, as well as more than 100 hours of robot trajectories and human-robot alignment data.
This path first lets the model accumulate a shared experience of how humans interact with the world, and then lets it learn through robot data how this body turns understanding into action.
On the execution side, Motus2 also uses an independent lightweight tactile expert to refine the short actions about to be executed and predict force feedback after contact. Vision is good at seeing where objects are and roughly how the world changes, while touch fills in the hardest "last centimeter" when contact occurs, giving the model an extra layer of real-time correction for delicate operations such as pressing, touching, grasping, and rotating beyond "being able to see."
Motus2 is still in the stage of extending from L3 to L4. Current validation is mainly concentrated on specific robots and manipulation tasks, and autonomous goal formation, persistent memory, open-environment generalization, long-term online learning, and safety control remain to be explored.
Its main progress is placing policy generation, future simulation, value evaluation, and tactile feedback into the same system, enabling robots to rehearse consequences before acting and use results to improve strategies after acting, providing a testable path toward L4 autonomous world agents.
From probabilistic learning to world intelligence
A Chinese technological path
The framework of general world models also continues the long-term research direction accumulated by Zhu Jun and Tsinghua TSAIL.
Professor Zhu Jun has long studied Bayesian methods, probabilistic machine learning, generative models, and reinforcement learning, respectively dealing with uncertainty, data distribution, future generation, and action decision-making.
TSAIL has also trained and influenced a group of researchers who have entered the world's leading teams. Song Yang collaborated with Zhu Jun on Bayesian learning research during his undergraduate years, and Song Jiaming also participated in related work during his undergraduate period, later producing important impacts in score-based generative models and DDIM directions respectively.
From probabilistic modeling to generation and decision-making, this academic lineage also reflects two types of entry points for AGI development. One starts from language and gradually enhances reasoning, tool calling, and task execution capabilities; the other starts from vision and the dynamic world and continues toward real-time interaction and embodied action.
Within China's AGI industry landscape, Zhipu and Shengshu can be seen as representatives of the two paths. Zhipu starts from language models and continuously enhances reasoning, Coding, and Agent capabilities; Shengshu starts from video generation and further enters real-time interaction, embodied action, and policy optimization.
The two routes are penetrating each other. Language Agents need visual feedback and environment models, and world models also need language to express goals, rules, and plans. Different entry points ultimately all point to the joint modeling of language, world state, and action.
Conclusion
Competition in world models will ultimately shift from the shock factor of Demos to the ability of systems to withstand feedback.
Images can create a credible first impression, geometry and physics determine whether this world can be used repeatedly, and action brings the model's judgment into reality. Whether a prediction holds must wait until after intervention occurs to be confirmed; whether a task succeeds must also be tested in unfamiliar scenarios, long-term operation, and continuous change.
The paper by Zhu Jun's team does not write the final answer to the world model debate, but it provides a set of questions that can continue to be tested: what the model understands, how far it can imagine, and whether it can correct itself after acting.
Over the past decade, the most important progress in AI has come from learning patterns in language and data. In the next stage, models will also have to face continuously changing environments. When models can continuously form judgments, bear the consequences of actions, and keep learning in such environments, world models will gradually move from a crowded technical term to a foundation on the path to general intelligence.
This article comes from the WeChat public account "机器之心" (ID: almosthuman2014), author: Yang Wen
