Gasgoo Munich- For the past few years, AI's landmark breakthroughs have played out almost entirely in the digital realm: writing text, coding, and generating images and video.
These capabilities share a common premise: the corresponding knowledge has long been sedimented on the internet in the form of text, images, and video, so models simply need to "read widely."
Robots, however, face a physical world governed by gravity and friction—one where slipping and dropping are constant risks. The tactile sense, force, and on-the-spot reactions required to get a job done have never been recorded at scale.
To bridge this chasm, "world models" have been thrust into the spotlight, widely regarded as the next major battleground for AI following large language models.
Yet the hotter the hype, the clearer the need for perspective: the ultimate goal of a world model isn't just to simulate a realistic world, but to understand and act upon the real one. Bridging that gap requires overcoming two major hurdles—physical intelligence and high-quality data—alongside "hard constraints" like power consumption and reliability.
Why World Models Are All the Rage
To answer that, we first need to see where the previous mainstream approach, VLA, hit a wall.
The VLA approach is essentially imitation learning: by observing countless samples of "what scene corresponds to what action," it learns to map visuals directly to movements.
In static, controlled environments, this method works well—grabbing a water bottle or moving objects across a desk is no problem. But VLA generalization depends heavily on training distribution: it scores near-perfect marks within known scenarios, yet performance degrades sharply once it steps outside that distribution—facing unseen objects or unfamiliar scenes.
World models aim to solve precisely this problem of "freezing up when things change."
Image source: ACE ROBOTICS
"Compared to VLA, the defining feature of a world model is its ability to generate and predict the future state of the world," said Wang Xiaogang, chairman of ACE ROBOTICS and co-founder and executive director of Sensetime, recently at the 2026 World Robot Conference forum.
More specifically, VLA excels at predicting motion trajectories in unchanging environments. World models, by contrast, can execute actions while simultaneously reasoning about and simulating various potential consequences in parallel spaces—ultimately integrating understanding, generation, and prediction.
To put it in a nutshell: VLA is "monkey see, monkey do," while world models are "look before you leap."
However, the consensus on the importance of world models doesn't mean VLA is obsolete. After more than a year of debate over technical routes, "fusion" has become the mainstream direction at the model architecture level. Companies like AgiBot, AI² Robotics, and Galbot are already exploring this path.
Yet there is still no standard answer for how, exactly, to fuse architectures internally or who takes precedence. What's more, the specific capabilities required for a world model to truly serve humans remain an open question.
In Wang Xiaogang's view, "three types of intelligence" are indispensable: generative intelligence, physical intelligence, and cognitive intelligence.
"Generative intelligence" refers to simulating and predicting the future world; "physical intelligence" corresponds to spatial intelligence—including spatial perception, memory, and mastery of physical properties and laws; "cognitive intelligence" involves breaking down long-horizon tasks and judging execution status.
Wang believes the most critical shortage right now is "physical intelligence," particularly the ability to understand spatial physical laws and causality. "Much of the video used to train models today isn't real footage—it's movie special effects and sci-fi shots." These visuals are inherently anti-physical and surreal, so naturally, they don't teach the laws of the real world.
Compounding this is a shortage of first-person interactive video, along with incomplete information on camera pose, 3D geometry, and multi-perspective views—all of which constrain the development of "physical intelligence."
After all, pretty visuals can fool the eye, but they can't fool Newton.
Beyond the "three intelligences," Wang notes a "harder" constraint often overlooked: current intelligence relies on stacking data, parameters, and compute. Robots, however, run on batteries and cannot recharge while operating like electric vehicles, creating a "ceiling" for on-device compute and power consumption. "Therefore, balancing the upper limits of intelligence with the low-power constraints of deployment is an urgent problem to solve."
Zhang Yufeng, founder and CEO of ANYVERSE DYNAMICS, agrees that for world models to truly serve humans, they must first grasp the underlying logic of the physical world, complete long-horizon tasks, and generalize across scenarios.
At the same time, he points out that just as humans don't need to master every detail of the physical world to complete complex tasks using an approximate "world model," embodied AI shouldn't strive for pixel-perfect replication of every detail. Instead, it should learn the high-dimensional representations and causal relationships that truly influence decision-making and action.
Zhang also emphasizes two easily overlooked points: safety and interaction. In an era of human-machine coexistence, relying solely on a large model isn't enough; there must be additional safety redundancy mechanisms. For interaction, robots need to read human demonstrations just as a child learns from an adult's example.
Hu Luhui, founder and CEO of Zhicheng AI, also pins the key on "understanding" and "causality": understanding physics and dynamics, planning tasks, and using causal reasoning to decide the next move. He adds that once robots enter homes, issues of safety, privacy, and ethics aren't just regulatory matters—they must be solved by the technology itself.
In industrial To B and To G scenarios, however, customers prioritize determinism, reliability, accuracy, and stable long-sequence output. Song Bin, co-founder and general manager of Feidu Technology, points out a series of unresolved challenges behind this: long-sequence stability under multi-physics coupling, spatial structure parsing in open environments, spatiotemporal semantic parsing, and on-device compute constraints and cost ROI—all require continuous industry effort to overcome.
In this sense, building a world model capable of "working in the trenches" is really a comprehensive exam: physical understanding and causality, long-horizon tasks and generalization, determinism and reliability, safety and ethics—flunk any one, and you fail.
And ultimately, the toughest question on this comprehensive exam points to one thing: data.
The Unavoidable "Data Barrier"
World models need data to bridge the physical intelligence gap—there's no way around it.
The reality, however, is that historical accumulation of the data most needed for embodied intelligence—"real humans operating in the real physical world"—is virtually zero. According to the China Academy of Information and Communications Technology (CAICT), currently available global high-quality real data amounts to just 100,000 to 1 million hours. For embodied foundation models to reach their "ChatGPT moment," at least 10 million hours are needed—a gap exceeding 99%.
Image source: Maniformer
But the size of the gap is just the surface. Digging deeper, the real bottlenecks at the data level are distributed across three stages.
First, collection methods are flawed; you can't capture real skills through "staging."
For a time, hiring specialists to collect data based on scripts was the industry standard. But in Zhang Yufeng's view, this path's effectiveness has been overestimated.
"We ran experiments ourselves: building sets, hiring collectors, pre-writing task scripts, and having them repeat prescribed movements. The improvement in final training results was very limited—frankly, poor—because the learning pattern was too singular," Zhang said.
By comparison, he favors "in the wild" collection—gathering data naturally within real production and life settings, the more natural the better. "For example, deep cooperation with manufacturing, retail, and hospitality sectors gives access to thousands of frontline workers, allowing data to accumulate rapidly. Through practice, we're already starting to see the scaling law at work."
Second, the collected data elements are incomplete.
Collecting manipulation data is far from simple as "shooting a video." Multimodal data—vision, force, and touch—is indispensable. Without it, the model has no way to judge how much force to apply during a grasp or whether an object is slipping.
"But through partnerships with many glove companies, we've found that current tactile gloves aren't mature enough," said Huang Yuanhao, founder and chairman of Orbbec. "Then there's synchronization—data captured by the eyes is from this moment, but force data from the hand might be 15 milliseconds off. That's already wrong."
Yet emphasizing quality doesn't mean blindly pursuing high precision.
"People usually assume the higher the precision, the better. But higher precision inevitably drives up costs. The core question is: at what level of data precision is it 'good enough' for the model? Only by identifying that threshold can we better balance data quality and scale," argues Shen Yujun, chief scientist at Robbyant.
More easily overlooked than precision is data distribution. In his view, "100,000 homogeneous tasks aren't very meaningful." Task coverage matters more than simply stacking numbers.
Third, the data cannot be reused; no matter how much you collect, it remains an "island."
A more subtle waste than collection is the lack of standards. Because robot configurations vary—degrees of freedom, link lengths, sensor layouts—data formats are inconsistent. This makes the vast amounts of data collected today naturally difficult to migrate across different robot bodies or companies, failing to alleviate shortages and instead driving up training costs.
Zhu Zheng, co-founder and chief scientist of Gigaai, crunched the numbers: training a world model and embodied foundation model with billions of parameters using several hundred thousand hours of data costs about 100 million yuan. If data volumes rise to tens of millions or even hundreds of millions of hours, and model parameters expand ten- to a hundred-fold without improving sample utilization efficiency, training a giant foundation model could cost over 100 billion yuan—more expensive than today's LLMs. That is "clearly unsustainable."
For this reason, he advocates collecting new data and improving quality while simultaneously iterating technology to boost sample utilization efficiency. Only then can models be trained within a finite budget.
Huang Yuanhao goes even further, arguing that while data is important, data standards are more important.
After all, without unified standards, the more data you collect, the more it looks like building towers on separate islands—no matter how high the towers, they won't connect. The value of standards is to transform data from "private assets" into "circulating industry infrastructure," allowing the accumulation of millions of hours to truly stack up rather than being wasted in duplication.
Crucially, only when data can be reused across different robot bodies and models will data costs come down. Otherwise, burning cash will be a bottomless pit.
Image source: ANYVERSE DYNAMICS
Consensus on Data Structure, Divergence on Training
With the bottlenecks identified, the next step is naturally how to break through them.
First, improve quality at the data source.
Zhang Yufeng points out that for world models, while data scale matters, data quality is even more critical. So, how do we judge if data has quality?
He offers a ruler: the "effective information density" of interactive data, categorizing data by value from low to high. The first category is observational data, with almost no real contact or effective action trajectories. The second is simple contact data, like grabbing or placing items. The third involves dense contact points and fine operation, like handling tiny or fragile objects—far more valuable. The fourth, often overlooked but extremely valuable, is data recording "failure processes."
"Especially when the scale and quality of this type of data are improved and mixed into training, it helps significantly increase success rates and assists models in recovering from reversible failure scenarios," Zhang said.
Song Bin suggests collecting and organizing data guided by physical laws, common sense, and spatial constraints. This ensures data matches training goals and saves substantial costs.
He also believes we should leverage existing industrial observation data systems—not just video and photos, but environmental field data like sound, light, and electromagnetism. By mining physical laws and organizing spatiotemporal topological relationships from this, we can push data from "correlation" toward "causality," truly establishing causal relationships based on the robot body itself.
Next, boost efficiency at the model training end.
Low efficiency doesn't just mean more severe waste and bigger gaps—it directly drives up training costs.
But an easily overlooked fact is that while embodied intelligence has a massive appetite for real data, world models will fundamentally alter this landscape. As the model's understanding improves, its dependence on data volume will drop sharply.
"When AI develops to the level of cognitive understanding, its dependence on data isn't as massive as people imagine—requiring endless redundant data. Humans learn a skill through understanding, not infinite data. This is the essential difference from supervised learning and conversational learning," Hu Luhui said.
Moreover, since world models possess data generation capabilities themselves, Hu believes scarce scenario data that is hard to collect can be simulated through the model's understanding.
Zhang Yufeng offers a rough estimate: the world model approach may rely on about two-thirds less data than pure imitation learning.
Finally, get the data flywheel spinning in real-world deployments.
Models must eventually leave the lab and deploy into real scenarios to continuously stream back data and iterate—only then does the positive flywheel spin. Otherwise, collecting data without actually using it is meaningless. So, how do we do it?
Zhang Yufeng proposes a path: "Industrial for skills, commercial for generalization, home last." First, hone basic skills in flexible processes where traditional automation falls short. Next, enter chain commercial scenarios—"similar in style but different in detail"—to train generalization capabilities. Only then break into the home environment.
Notably, this "easy-to-hard" penetration sequence is a pragmatic path recognized by many in the industry.
But even if these directions form a consensus, not all problems are solved. On a crucial issue, the industry still remains divided: in the post-training and fine-tuning stages, should we rely primarily on real robot data or synthetic simulation?
Image source: DexForce
Jia Kui, at DexForce, has consistently insisted on using synthetic data and generative simulation for post-training. His reasoning is pragmatic: if costs are forever burdened by expensive collection and debugging, the business model won't hold up.
"So, starting from first principles, we must pursue a more efficient way to produce post-training data. Generative simulation and synthetic data are critical paths to solving the physical generalization problem," Jia noted.
The mainstream, by contrast, leans more toward "real robot anchoring."
Wang Xingxing, chairman, general manager, and CTO of Unitree, has repeatedly pointed out in public that true AI robot training data requires massive amounts of human or internet data as primary pre-training data, supplemented by a portion of real robot data, to truly match the robot with the physical world.
Behind this lies the "reality gap" from simulation to reality (sim2real): simulators struggle to fully replicate the friction, deformation, motor delays, and sensor noise of the real world. When strategies migrate to actual robots, varying degrees of performance degradation occur.
To put it in perspective, it's like practicing driving in a game—you might master the mechanics, but once you're on the real road, you still have to readapt to the steering resistance and tire grip.
Regardless of the route dispute, one undeniable fact remains: the data barrier cannot be overcome by a single company alone. After all, the data foundation is a public good for the industry. Technical routes may diverge, but standards and public data are the road the entire industry must build together.
Conclusion
The industry consensus on world models is clear: they are the hope for breaking through the generalization bottleneck in embodied intelligence; physical intelligence is a shared weakness; the data flywheel determines iteration speed; and capabilities must ultimately be forged in real-world deployments, not in the laboratory.
The divergences, however, are equally real: technical routes have not yet converged; whether post-training should rely on real robots or synthesis; how to balance various data types; and whether world models are fundamentally a "generation problem" or an "understanding problem"—all remain unresolved.
As for the timeline, keep two coordinates in mind: Wang Xingxing's "Double 80%"—completing 80% of tasks in 80% of unfamiliar scenarios, achievable in as soon as two to three years, or as long as five to ten; and Zhang Yufeng's "Industrial-Commercial-Home" penetration sequence. One provides a quantified finish line, the other a realistic path. Put them together, and you have roughly where this technological long march stands today.
World models have already learned to imagine the future in their "minds." What they still need to learn is how to bow down and, using a pair of hands trained on high-quality data, get the work in the real world done one task at a time.
The distance from "imagining" to "doing" will be the hardest kilometer yet for world models.
