In late September, AMD acquired World Labs, co-founded by Li Feifei, in an all-stock deal worth USD 8.2 billion. This is the second-largest acquisition in AMD's history, bringing world models into the industrial spotlight in such an expensive manner for the first time. But the price only answers half the question: how much a company is willing to pay for this direction and whether this technology can actually work in factories are two different things.
A Chinese company, ShadowAI, has given a different answer: bringing the 4D world model into shoe-making workshops first.
01. From 3D to 4D: Why the Extra Dimension is Valuable
Currently, the growth and evolution speed of AI in the digital world is beyond imagination, yet in the real physical world, technological evolution is significantly lagging behind. The so-called native 4D world model superimposes the time dimension on top of the 3D space and incorporates mechanical and topological information: what it needs to capture is not only what objects look like, but also how they deform, how they interact, and how they evolve over time.
Most current approaches are stuck in the middle: 2D video generation has the visuals but lacks real depth; static 3D reconstruction has depth but lacks time. Missing this dimension leads to errors in reasoning that violate common physical sense, such as objects penetrating each other or gravity failing out of nowhere.
World Labs and ShadowAI set off almost simultaneously in 2024, and their judgments on technology are highly consistent: for AI to move towards physical intelligence, it must first establish an understanding of 4D space. The real difference lies later—World Labs chooses to build a virtual world first and then put agents into it; ShadowAI chooses to step into factories first.
What can this route deliver? ShadowAI's self-developed 4D foundation world model "Pojie S1" provides a set of quantifiable answers. It integrates rendering, simulation, and planning into the same foundation. The improvements brought by unified representation are quantifiable: the success rate of image goal navigation increases from 43% to 72%, language goal navigation reaches 69%, and the inference speed is about 45% faster than the previous best method; on nine dynamic manipulation tasks, the average success rate is more than 10 percentage points higher than the previous best method. Sample efficiency is even more noteworthy: with only 15 real demonstration data pieces, the model can approach peak performance.
02. ShadowAI: From Tsinghua Laboratory to Shoe-Making Workshops
Founded in 2024, ShadowAI focuses on native 4D data, world models, and embodied intelligence, with two main directions: physical embodiment and digital entertainment. The technology stack consists of three layers: 4D data collection, 4D foundation world model, and vertical domain models. Currently, this system is already running in real industrial scenarios and has secured orders.
Most teams working on world models only focus on algorithms. Since its inception, ShadowAI has put scientific research, engineering, and business on the same table.
Founder and CEO Min Wei holds a Ph.D. in Engineering from Tsinghua University and is a Senior Engineer. He formerly served as the technical lead for the Alibaba Local Services robotics team, possessing over 20 years of experience in AI and robotics R&D and deployment; Chief Scientist Liu Yebin is a Tenured Professor in the Department of Automation at Tsinghua University and a recipient of the National Science Fund for Distinguished Young Scholars. His research focuses on 3D vision, 3D and 4D content generation, and digital humans, with multiple achievements having completed industrial technology licensing; Special Science and Technology Advisor Zhu Yu is a Tenured Professor in the Department of Mechanical Engineering at Tsinghua University.
The academic background enables it to keep up with the frontier, while industrial experience lets it know how many constraints exist on-site. This is also the reason why ShadowAI did not take the laboratory route but chose real2real: turning the 4D world model into a capability that can be continuously iterated on the production line.
03. Finding Data on the Internet or on the Production Line
In a late July podcast by a16z, Li Feifei proposed a judgment: what is holding back embodied intelligence is not computing power, but the "data desert," which is also recognized as the biggest bottleneck.
World Labs' solution is Real2Sim2Real: moving the real environment into the simulation space, letting agents make a large amount of trial and error inside, and then deploying them back to the real world. In terms of data, it mixes publicly available internet videos with proprietary materials, aiming to scale up the data volume at a low cost.
Min Wei's another layer of judgment is: in embodied intelligence, data has never been a question of quantity, but of dimension. Internet videos mostly record standard, successful operations, and rarely record the frequent mistakes, anomalies, and material deformations that occur on production lines. But factories are precisely full of uncertainties—it is these "non-standard" samples that can truly teach robots how to handle complex working conditions.
Therefore, the core of ShadowAI is "data rehearsal > collection," adhering to real2real, directly collecting native 4D data 24/7 on shoe-making production lines, turning workstations into training grounds, and feeding data back to the 4D foundation model while the production line is running.
However, there is an issue that cannot be avoided—high-quality 4D data has always been expensive. Traditional multi-view dynamic collection solutions rely on camera arrays with dense viewpoints, and the entire set of equipment often costs several million CNY, making it difficult to popularize 4D data. To bring it into the industry, the cost at the collection end must first be reduced.
ShadowAI's solution is the self-developed "ShadowAI 360" multimodal collection system: requiring only 4 to 6 consumer-grade RGB cameras paired with ordinary graphics cards, it can achieve dynamic collection at 25 frames per second, producing 4D data with millimeter-level precision in real time. Compared with traditional solutions, the comprehensive hardware cost is reduced by over 90%.
04. From One Robot to a Production Line
Hardware is just the entry point; for data to become capability, there must be a training ground. In September this year, the Embodied Intelligence Training Center in Shougang Park, Shijingshan District, was unveiled. This is the country's first embodied intelligence data and training infrastructure driven by a 4D world model. ShadowAI is the core co-builder and the main builder of the first-phase 4D foundation world model. The center has iterated from 1.0 to 2.0, building a data flywheel of "collection-modeling-training-validation-feedback," integrating four types of complementary data: real humans, real machines, 4D, and simulation. Using the 4D world model as a unified foundation, it connects three types of intelligence: digital intelligence, spatial intelligence, and physical intelligence, linking the entire chain from cognitive decision-making and scene reconstruction to real machine execution. Therefore, the model can continuously consume first-hand data from real production lines.
Whether the flywheel spins or not ultimately depends on the production line. Facing flexible manufacturing, ShadowAI has currently developed three robots: ShadowGlue for glue application, ShadowPress for sole pressing, and ShadowBox for packaging. They target the long-standing "dual flexibility" challenge in the shoe-making industry—leather and fabrics deform under force, which is material flexibility; rapid style iterations and non-standard incoming materials, which is task flexibility. It is precisely these two points that have kept the penetration rate of traditional rigid automation equipment stagnating at 5% for a long time.
The ShadowGlue glue application robot can autonomously learn process logic simply by watching workers' operation videos. It can perceive changes in glue viscosity in real time and complete motion compensation within 0.1 seconds, reducing the material waste rate to less than 1%; the ShadowPress sole pressing robot achieves sub-millimeter precise alignment through dynamic compensation based on visual algorithms, reducing the end-effector precision requirement by 10 times. Its yield rate is stable at over 99.9%, and it can operate unmanned 24/7.
The ShadowBox packaging robot was successfully exhibited at the booth of the 23rd China-ASEAN Expo in 2026 and the 5th Jingxi Regional Development Forum. Relying on the 4D world model, it identifies changes in the shape and pose of paper boxes in real time and uses dual robotic arms to collaboratively complete flexible paper box folding.
In the first half of 2026, ShadowAI reached cooperation with several leading Vietnamese shoe-making enterprises such as Baolong Xieye, Shunfei Xieye, and Yuehan Xieye, planning to deliver hundreds of ShadowGlue glue application robots and ShadowPress sole pressing robots, with an order scale of tens of millions of CNY. As of the closing of the China-ASEAN Expo, the cumulative signed orders for ShadowAI's embodied products have reached hundreds of millions of CNY, and business in the ASEAN region is still expanding.
Behind these numbers lies a layer that is even harder to replicate. The barrier for ShadowAI is not in any single algorithm, but in the stacking of three things: the collection cost that allows operation with just 4 to 6 cameras, which frees the data scale from being constrained by equipment budgets; the real data fed back 24/7 on the production line, which allows the model to obtain samples from the real physical world; and the pathway from collection to feedback at the Shijingshan Big Data Training Center. Only when these three layers are combined can the flywheel spin; missing any one layer, the 4D world model can only stay in demonstrations.
Final Words
Capital prices the technology direction, and factories score the technology based on cost reduction and returns. The significance of ShadowAI goes far beyond being the "Chinese version of World Labs."
The real test lies in whether this complete chain from 4D data and 4D world models to robot execution can be extended to more factories and more task scenarios.
This major exam has just begun.
This article is compiled based on publicly available network information, for reference only, and does not constitute investment advice.