The Point Is Realism
By Jia Haonan, reporting from Aofeisi Temple
Would you believe it? The most engaging, authentic, and “hands-on” booth at this year’s WRC was actually a coffee shop.
——Its style was completely different from the robot booths with restricted areas and isolated demonstration zones:
A robotic-arm “barista” inside the counter was hand-brewing coffee:
Next to it, a human barista wearing data-collection equipment was making latte art:
Outside the counter, a robot carried coffee cups through the crowd. After delivering one order, it would casually collect the cups from the neighboring table. When turning a corner, it even steered around a visitor who was taking photos:
The entire setup ran openly, with customers free to move around. Every now and then, a mischievous customer would even deliberately block the coffee cup the robot was preparing to set down.
There was no scripted performance, no fixed route, and no specially modified environment. People kept coming in to order coffee. This was a completely real coffee shop.
Only after asking around did we learn that this setup had already been operating on a trial basis at a high-end coffee shop for some time. It was brought to the WRC venue only after its stability and safety reached the expected level.
Among the hundreds of robot booths at WRC, most demonstrated, “I can perform this motion.” This one demonstrated, “I can keep working in a real environment like this without anything going wrong.”
We’ve seen plenty of robots hand-brewing coffee. What’s different this year?
The biggest difference you notice at first glance is not the coffee-shop cosplay, but the fact that this is a real shop operating inside the booth—
HOLLYS is not a fictional sign invented by a robotics company for the exhibition. It is a genuine, long-established Korean coffee chain:
The company has actually opened a co-branded café with Wujie Dynamics at the exhibition center in Beijing’s Yizhuang district. The only difference is that all the servers are robots:
The biggest attraction was the robot waiter moving among the customers. It uses a wheeled base and a semi-humanoid design, and wears an apron and an adorable hat.
Inside the counter, the robotic-arm barista prepared hand-brewed coffee for customers to sample. On the other side, human coffee masters prepared various drinks based on customer orders, creating beautiful patterns in the milk foam. Once a latte was ready, it was placed in the pickup area, and the robot waiter outside the counter immediately set off. It picked up the coffee cup steadily, planned a route through the open, crowded area, navigated around visitors taking photos, and stopped beside a table.
A camera sat in the upper-left corner of the tabletop. After recognizing it as a personal belonging, the robot avoided it and placed the coffee in an empty spot on the table.
A moment later, the customer left, leaving behind an empty cup and a wad of used tissues.
The robot picked up a basket and returned to the table. It steadily grasped the empty cup, then gently pinched the tissues and placed them in the basket:
The robot waiter then walked to the collection area and emptied the food waste into a trash bin. After that, it actually reached under the sanitizing lamp and disinfected its own hands:
Throughout the process, there was no preprogrammed route. Customers could move around at any time, and the tasks changed randomly.
△ Kim Jin-dong, Economic Minister, and Shin Sang-ryeol, Minister-Counselor for Science and Technology at the Embassy of the Republic of Korea in China, visited the Wujie Dynamics booth in person to try out the system and give it a thumbs-up.
Every decision—where to go, where to place something, what to collect, and which task to perform first—was made on the spot.
A human barista and waiter could complete this entire sequence without much effort. For a robot, however, it involves:
Dynamic environmental perception, real-time path replanning, semantic understanding—distinguishing personal belongings from trash—soft-object manipulation, such as handling tissues, long-horizon task planning, including carrying, clearing, collecting, and returning, as well as coordinated interaction with other robots, agents, and people in the environment.
Beyond long-horizon, continuous tasks in complex environments, the coordinated operation of multiple robots under fixed boundary conditions was also impressive, as demonstrated by the “mini production line” outside the café:
It could also complete a more delicate bead-threading challenge. This indicates that it already possesses the foundational capabilities required for multi-step coordinated tasks and precision manufacturing:
Each challenge represents an independent frontier research topic in robotics.
Being able to continuously complete this entire sequence of actions in an open space was itself one of the biggest differences between this system and most of the projects on display at the WRC venue.
As the exhibition drew to a close, the booth staff were tired, and so was the human barista inside the counter. But the robot waiter, Xiao K, which had also been working nonstop for five days, continued to operate smoothly.
The goal is to have robots gradually take on repetitive, demanding, and even dangerous work, using the same general-purpose brain to reliably complete multiple different tasks.
That is the true value of embodied intelligence: working reliably in the real world, creating value, and ultimately serving people.
Who built it? How does it work?
Wujie Dynamics was founded in Beijing in March 2025 and has continued to receive strong recognition from leading industrial and financial investors worldwide. In the first half of this year alone, it completed several rounds of financing totaling several hundred million US dollars.
At the product level, Wujie Dynamics is pursuing a full-stack self-development strategy covering the “brain, body, and core components.” The K15 robot passed the EU’s industrial-grade CE certification covering all applicable directives in July 2026, becoming the world’s first embodied-intelligence robot to receive industrial-grade CE certification.
In terms of commercial orders, the company has publicly announced a cumulative total of RMB 700 million, with the first batch of products already shipped to Europe.
The founding team all comes from the autonomous-driving industry.
Founder and CEO Zhang Yufeng joined Horizon Robotics in 2017. He previously served as vice president and president of its Intelligent Vehicles Business Unit, leading Horizon’s advanced-driving systems business to become the market-share leader among Chinese domestic passenger-vehicle brands. He also spearheaded the mass-production deployment of Horizon Journey chips in the ADAS field and led multiple strategic partnerships with major domestic and international companies, including BYD, Changan Automobile, Volkswagen Group, and Continental.
Co-founder and CTO Xia Zhongpu holds a PhD from the Institute of Automation of the Chinese Academy of Sciences. He previously led the prediction module at Baidu Apollo and later served as head of end-to-end technology at Li Auto. Under his leadership, the company took the end-to-end solution from demo to mass-production delivery in fewer than 100 days, making Li Auto the first automaker in China to roll out an end-to-end urban NOA solution across its entire fleet.
The two founders share a common experience: both went through the complete transition in autonomous driving from rule-based systems to end-to-end approaches.
The key insight they accumulated during this process, and transferred to embodied intelligence, is that simply piling up data cannot solve the generalization problem in open environments.
This judgment directly shaped Wujie Dynamics’ view of the VLA approach currently popular in autonomous driving.
Zhang Yufeng put it bluntly: VLA is essentially still imitation learning. “It doesn’t drive over the curb, not because it understands what a curb is or what would happen if it drove over one, but because it has seen a great many curbs in different scenarios, along with data showing vehicles not driving over them.”
The autonomous-driving industry has already demonstrated one thing: data accumulation can make a system perform very well in scenarios it has seen. But the environments that embodied intelligence will ultimately enter—whether factories or homes—are highly open and impossible to enumerate exhaustively. If the goal is merely to have robots “memorize” more scenarios, the approach will not get very far.
At present, the industry is clearly divided over how to build a world model: pixel-level generation versus latent-space prediction.
The representative of the pixel-level generation camp is the video-generative world model, which attempts to predict what the pixels in the next frame will look like. The image must be realistic, the lighting and shadows accurate, and the textures clear.
Wujie Dynamics, by contrast, has chosen a latent-space world model. Its central idea is to have the model understand “how physical changes occur,” rather than “what action should correspond to a particular image.”
This was also the point Zhang Yufeng repeatedly emphasized during the main forum at this year’s conference:
At the WRC main forum, the “World Model Expert Committee of the Chinese Institute of Electronics” was formally established. Zhang Yufeng, founder and CEO of Wujie Dynamics, was invited to serve as a committee member and took part extensively in the subsequent dialogue, titled “The Real-World Distance from World Models to Serving Humanity.”
“Humans do not need to master every detail of the physical world to complete complex tasks using a roughly accurate ‘world model.’ Embodied intelligence likewise does not need to pursue a pixel-perfect reconstruction of every detail. Instead, it should learn the high-dimensional representations and causal relationships that truly influence decisions and actions.”
The approach is to reduce reliance on pixel-level computation and on massive amounts of data tied to fixed tasks and specific scenarios. Reinforcement learning can then be used for trial and error, feedback, and rewards, enabling robots to develop not only a “worldview” that helps them understand the world, but also a “value system” for judging the value of different actions.
In June 2026, Wujie Dynamics released MWA™, its general-purpose embodied-intelligence brain and latent-space world model. It ranked first worldwide on the RoboCasa GR1 TableTop leaderboard launched by Stanford University and other institutions, achieving an average task-success rate of 75.2%—ahead of mainstream models such as NVIDIA GR00T-N1.6 and XPeng DIAL:
The connection between this leaderboard and the coffee-shop scenario is essentially two ways of validating the same underlying logic.
The leaderboard is characterized by a high degree of randomization: nonstandard kitchen environments, highly variable lighting and shadows, and interference from miscellaneous objects. It tests a model’s ability to handle change in a virtual environment.
The coffee-shop scenario tests the same ability to handle change: real customer traffic, unpredictable tabletop conditions, and changing light. For example, when someone reached out and blocked part of the table, the robot paused for a second, replanned its actions, and placed the cup in an empty spot nearby.
Similarly, when the robot folded boxes, folded clothes, or threaded beads, what the model learned were the patterns of change that occur during physical operations such as folding, aligning holes, and handling soft fabric.
When the bead would not pass through, for instance, the robot adjusted the angle and tried again. This was based on the model’s prediction of how the string would deform under force, rather than on matching pixels to a specific visual scene at a particular angle. It was therefore more like a person exploring a shape by feel than relying on visual memory.
But one question remains unanswered: what does this non-mainstream approach mean when measured against the three questions most frequently asked about embodied intelligence—commercialization, real-world deployment, and scaling?
Embodied Intelligence’s First Complete Chain of Evidence
Among the hundreds of companies at the WRC exhibition, Wujie Dynamics had a distinctive “presence” that was difficult to ignore.
Most of the more than one hundred exhibitors repeatedly demonstrated preset actions in controlled environments: fixed workstations, fixed lighting, fixed object positions, and fixed movement trajectories. What visitors saw was, “The robot can perform this specific action under these specific conditions.”
The real coffee shop, which operated for five days and received more than 1,000 visitors, represented more than popularity. It meant that the robot waiter, Xiao K, underwent an intensive, long-duration test of reality across three dimensions in an open commercial setting with a continuous flow of people: dynamic spaces, non-fixed task sequences, and real-time human-robot interaction.
In other words, the coffee-shop scenario was not an “exam question” specially designed by Wujie Dynamics for its robots. It was the natural application of MWA™’s training logic: concrete tasks in an open environment, making it an ideal setting for testing the ability to “understand physical change.”
For most embodied-intelligence companies, industrial scenarios are the starting point for commercialization. That is both natural and reasonable: tasks are fixed, environments are controllable, and customers have a strong willingness to pay.
But Wujie Dynamics has taken a deeper view. It has not treated industry as the sole destination, but has advanced industrial and commercial applications in parallel. On the industrial side, it develops skills, such as seat-belt operations for ZF and work at Envision Energy’s battery factory. On the commercial side, it develops generalization, as demonstrated by the coffee shop.
△ EDAILY, a well-known Korean financial media outlet, visited the exhibition and interviewed Zhang Yufeng, founder and CEO.
Behind this choice lies a specific judgment: industrial scenarios can verify whether a machine can perform operations stably and precisely, but they cannot fully verify whether it can continue making decisions in an open environment, because industrial environments are essentially designed to “eliminate uncertainty as much as possible.”
If the ultimate destination of embodied intelligence is the completely open environments of homes and the outdoors, then its generalization capabilities must be validated and iterated in more uncertain scenarios in advance, rather than waiting until robots enter the home to start from scratch.
The coffee shop sits precisely at the midpoint between “concrete tasks and open environments.” Its SKU range is relatively limited—coffee, cups, tissues, and chairs—but changes in people, lighting, and spatial positions are entirely uncontrolled. If a model can operate continuously and reliably in a coffee shop, it indicates that it has passed a basic test in the dimension of “continuous decision-making in an open environment.”
Industrial scenarios are responsible for refining operational skills, while commercial scenarios develop generalization capabilities. Both lines share the same general-purpose brain, MWA™. High-quality data accumulated in real-world scenarios continuously feeds back into the model, driving a virtuous upward spiral.
This is a mutually reinforcing structure, not two independent business directions.
Across the three levels of scenarios, technology, and business, everything points to the same first principle Wujie Dynamics is putting into practice: scenarios are infinite, but physical laws are limited. Identify which changes are relevant to a decision, understand how objects move after forces are applied, and replan when space becomes occupied—and with these capabilities, there is no need to enumerate every possible scenario.
The hardware and software of embodied intelligence are both still far from converging. On the hardware side, the industry is exploring everything from planetary and harmonic reducers to linear modules and dexterous-hand failure rates. On the software side, it is exploring everything from VLA and video-generative world models to latent-space world models. Every major approach is still under investigation; none has yet been conclusively disproven or proven.
Against this backdrop, assessing the substance and competitiveness of embodied-intelligence players cannot come down to whether they are “trendy.” The only meaningful criterion is whether their capabilities can be verified.
Wujie Dynamics has published research papers, achieved public results on simulation benchmarks, operated a real coffee shop on a trial basis, and secured commercial orders.
It offers a rare, verifiable chain of commercialization evidence amid the “big bubble” phase of embodied intelligence.