The Moat in Embodied Intelligence Is Not Shipment Volume
By Tian Yanlin, reporting from Aofeisi Temple
After walking around this year’s WRC, I found the simplest and most straightforward way to tell whether a humanoid robot company has really demonstrated its capabilities:
See whether it dares to bring real-world operating scenarios from its customers’ factories to the exhibition floor.
Promotional videos can be edited and demos can be rehearsed, but production lines will not cooperate with actors’ stage positions.
Dimensional tolerances and cycle-time pressure when loading and unloading automotive sheet-metal and machined parts, randomly stacked small parcels in logistics, positioning and movement during continuous operations—any detail could cause a robot to get stuck on the spot and trigger an exception mode.
UBTECH directly placed these challenges, which it had overcome in customer factories, on its exhibition stand at a 1:1 scale.
On site, we saw the industrial humanoid robots Cruzr Y1 continuously performing depalletizing and palletizing operations and transporting cartons and material bins. Cruzr S2, meanwhile, handled the loading and unloading of automotive machined and sheet-metal parts, completing submillimeter-level positioning on the workbench.
Elsewhere, Cruzr S2 picked skincare-product boxes of different specifications from mixed material bins and placed them one by one on Dema Technology’s automatic seeding wall packing station to complete seeding and sorting. A rough estimate puts its average picking throughput at nearly 1,100 items per hour.
With no human intervention and no preset programs, the robot autonomously handled everything from recognition, movement, and picking to placement.
Even when a pick failed, the robot could identify the problem autonomously, adjust its strategy, and try again.
Behind this lies a particular view from the “old hand” of humanoid robotics on the competition in 2026:
Customers pay for real-world scenarios.
Dr. Jiao Jichao, vice president of UBTECH and dean of its Embodied Intelligence and Humanoid Robotics Research Institute, told QbitAI that when customers select a robotics supplier, they do not primarily look at how well the company performs on a particular benchmark. They first look at whether the hardware is stable, whether the robotics company understands the scenario, and whether the solution can truly operate in practice.
Completing a task once in a laboratory may test the upper limit of an algorithm.
But when a robot enters a factory and performs thousands of operations continuously, the test involves more than algorithms. It also covers hardware reliability, task success rate, cycle-time efficiency, procurement and maintenance costs, and the robotics company’s understanding of the entire production process.
The solutions UBTECH showed at WRC have already moved beyond the practical training stage and entered the small-scale delivery phase. The company is currently expanding their scale based on their performance in customer scenarios.
The more operations it completes in real-world scenarios, the more UBTECH has also built its own ecosystem.
Jiao Jichao candidly acknowledged that once humanoid robots truly enter customer sites, it is difficult for a single company to cover the entire chain.
Chips, components, complete-machine manufacturing, scenario delivery, and data feedback—every link determines whether robots can progress from a few trial units to replication across hundreds or thousands of units.
The Exhibition Stand Turns into a Factory, Recreating Customer Production Lines at 1:1 Scale
At this year’s WRC, UBTECH did not bring only one star robot. Instead, it put all three of its product lines—industrial, commercial, and home consumer—on the exhibition floor at the same time.
On the industrial side, UBTECH recreated customers’ real production lines at a 1:1 scale. Nearly ten industrial humanoid robots, Cruzr S2 and Cruzr Y1, worked continuously at simple, repetitive, and monotonous stations involving automotive loading and unloading, transportation, depalletizing and palletizing, and sorting—from the 9 a.m. opening until the exhibition closed at 5 p.m.
On the commercial side, UBTECH brought the new commercial service humanoid robot Walker C1, designed primarily for reception and guided tours, cultural and entertainment performances, research and development, and intelligent teaching assistance.
For the home consumer market, the biomimetic humanoid robot U1, equipped with the “Resonance-LM” emotional foundation model, also made its debut. It focuses on emotional support and companionship.
Putting the three product lines together was not simply a matter of expanding UBTECH’s SKU lineup.
The most visually striking sight was not all the robots standing in a row, but Cruzr S2 and Cruzr Y1 directly entering a continuous, uninterrupted operating state.
With no human intervention, they autonomously completed the entire sequence of recognition → picking → movement → positioning → placement → inspection → return. The smoothness of this continuous workflow fully reflected the most realistic and demanding requirements of industrial scenarios:
High-intensity, long-duration, stable, continuous operation.
A closer look at the details reveals that this was by no means a carefully choreographed exhibition performance. It completely replicated the logic of customers’ daily production operations.
For example, loading and unloading automotive machined parts requires two robots to work in coordination and transport workpieces larger than one meter, while also handling variations in incoming materials, calibration drift, workpiece deformation, and on-site disturbances. The final positioning accuracy must be less than 1 mm.
In automotive sheet-metal loading and unloading scenarios, the robots must repeatedly perform loading, positioning, inspection, and unloading between the conveyor line and the machine tool. The VLA model runs on the edge, with positioning accuracy likewise below 1 mm.
Sorting small items for logistics and e-commerce requires the robot to continuously pick skincare-product cartons of different specifications, colors, and stacking conditions, with an average picking throughput approaching 1,100 items per hour.
Real factory operating scenarios impose entirely different capability requirements on humanoid robots.
The challenge is no longer limited to performing movements smoothly or making running and jumping look impressive. Robots must instead understand complex real-world environments, autonomously break down task objectives, reliably complete actions such as picking, assembly, and placement despite uncertain disturbances, and autonomously correct errors and retry after failures.
This creates a vast gap between task-oriented “worker robots” and performance-oriented “dancing robots” in hardware selection and computing configuration.
Dancing robots prioritize explosive power and agility, focusing on visual effects such as running, backflips, and dance stunts.
Their hardware tends to use lightweight designs. Most stand around 1.2 meters tall and generally are not equipped with visual or force sensors, nor do they require end-effectors such as dexterous hands.
Their joints typically use planetary gear reducers, which offer high flexibility at a controllable cost and are sufficient to support highly dynamic physical performances. On the computing side, chips such as Rockchip’s RK3588 are generally adequate. These chips primarily serve low-level motion control, executing only preprogrammed trajectories rather than running complex embodied foundation models.
By contrast, “worker robots” generally stand over 1.7 meters tall and are designed for real operating conditions involving sustained loads, long periods of continuous operation, and frequent physical contact and interaction with objects.
Their joints must withstand continuous loads while offering fatigue resistance and precise control of contact torque. They therefore use harmonic-drive reducers, which cost more but provide superior rigidity and stability.
At the same time, robots must perform environmental perception, force feedback, task decomposition and planning, and autonomous retries after failure. This requires running embodied foundation models and world models locally for real-time inference, making high-computing-power chips such as NVIDIA Thor essential as the computing foundation.
Simply put, dancing robots compete on dynamic performance, with their entire hardware configuration optimized for stage presentation;
worker robots compete on complete-machine reliability and autonomous intelligence. Their physical hardware and computing architecture are designed entirely around actual production operations, and the two hardware systems cannot be directly interchanged or reused.
By bringing such demanding real production lines directly to the WRC exhibition floor, UBTECH was clearly doing more than showcasing technology.
Hardware determines whether a robot can “move,” but the embodied-intelligence brain determines whether it can truly “get things done.”
So, on a crowded exhibition floor with no preset safeguards, what kind of brain is UBTECH using to support this high-intensity, uninterrupted battle?
Giving Robots an Embodied Brain
UBTECH’s “embodied brain” is essentially a three-layer system consisting of understanding, prediction, and execution.
The first layer is the Thinker foundation model, responsible for enabling robots to “understand” the world.
A robot must first determine where it is, what objects are in front of it, and what task the user has given it before it can plan and act.
Thinker uses first-person robot-view data to build its understanding of vision, language, and spatial environments.
△ UBTECH’s embodied-intelligence foundation model Thinker architecture diagram
According to publicly available information, Thinker ranked first in nine categories in authoritative benchmark evaluations of embodied-intelligence brain models under 10B parameters, while its largest foundation model reaches 100B parameters.
These evaluation results verify the model’s fundamental capabilities in perception and understanding. But once it enters production, it must also face a constantly changing physical world.
That requires the second layer: prediction.
Based on the Thinker foundation model, UBTECH developed the “Thinker-WM world model” to answer the question:
What will happen next if I do this?
Will an object fall after the robot picks it up? Will moving to a certain position cause a collision? Does the next action comply with the laws of physics?
To this end, Thinker-WM builds on Thinker’s accumulated capabilities in data pipelines, model debugging, and training infrastructure. It also upgrades the model architecture and introduces technologies such as Diffusion Transformer and Flow Matching, training video representations, action representations, and prediction within a unified representation space so that action planning better conforms to the laws of the physical world.
Thinker-WM currently ranks first on the LIBERO embodied-intelligence evaluation benchmark.
The final layer is Thinker-VLA, responsible for “execution.”
At the industrial workstations demonstrated at WRC, Thinker-VLA must convert task understanding into continuous control signals and determine, in real time, the robotic arm’s trajectory, end-effector orientation, grasping force, and placement position.
When a workpiece is displaced or a grasp fails, the model must also adjust the robot’s actions based on feedback so that it can continue completing the task.
VLA focuses more on manipulation and primarily answers, “How should I do this?” The world model focuses more on answering, “What will happen after I do this?” and making predictions in that regard.
Jiao Jichao said that only by connecting the three capabilities of understanding, prediction, and execution can robots move from “being able to see and think” to “being able to work.”
It is reported that UBTECH plans to explore integrating the three models into a unified end-to-end architecture, including backbone-network integration and parameter sharing.
But for robots that truly enter factories, model capability is only the first hurdle.
The next question is: Can this brain run in real time on the robot itself?
If every robot action depends on feedback from the cloud, network latency, communication stability, deployment costs, and power consumption will all become obstacles to industrial adoption.
UBTECH’s direction is clear: move highly real-time capabilities to the edge.
According to the company, after edge-side optimization, Thinker-VLA’s inference efficiency improved by 176%, storage usage fell by 60%, and the GPU memory requirement for all edge-side functional modules dropped from 64 GB to 32 GB.
Jiao Jichao said the team has migrated a large number of algorithms from the x86 platform to the low-power ARM platform while optimizing the underlying software, operators, and domain controllers. The goal is to run the entire software system on a single embedded board.
The more circuit boards you install, the more space, battery life, and reliability become issues. That is why we have always insisted on processing all software and algorithms on the edge.
In his view, the cloud can continue to handle slow inference, multi-robot scheduling, and model management, while the edge is responsible for the robot’s real-time perception and motion execution.
This means that models, computing power, power consumption, communications, motion control, and hardware reliability must be considered collaboratively from the earliest stages of product design.
Edge deployment solves the question of whether the embodied brain can be fitted into the robot’s body. But then, how can the same “brain” be quickly adapted to industrial, commercial, and home robots?
UBTECH’s answer is “1+N.”
“1” is a shared technology foundation, while “N” represents robot products defined for different scenarios.
Jiao Jichao said that until embodied technology achieves true generalization, products can only be defined in reverse according to different scenarios.
Industrial robots focus on payload, precision, and continuous-operation capabilities; commercial robots focus on high dynamics and interaction; and home robots need to understand facial expressions, emotions, and memory.
The same intelligent foundation can be reused, but the bodies of different robots still need to be engineered and optimized around specific scenarios.
The value of UBTECH’s ability to reuse technologies is already becoming apparent.
The commercial service humanoid robot Walker C1 was developed on the basis of existing industrial robotics technology. It reuses part of the software system and UBTECH’s self-developed servo technology, and took approximately four to five months from product definition to prototype completion.
Previously, developing a similar product entirely from scratch would typically have taken six to nine months.
UBTECH Shows Off Its “Circle of Friends”
Putting the embodied brain into the body and deploying it at the edge solves the first step in enabling robots to “do the work.”
In the end, competition in embodied intelligence will inevitably return to data.
For traditional AI, model capability depends to a large extent on the scale of internet data. For embodied intelligence, however, robots face a physical world full of uncertainty.
The same grasping action may produce entirely different results due to differences in object weight, changes in placement, and accumulated mechanical errors.
Such feedback from real environments is difficult to obtain through simulation alone.
Therefore, robots must enter real-world scenarios and accumulate experience through repeated task execution, failure correction, and interaction with the environment.
According to reports, real-robot data accounts for as much as 60%–70% of the data used to train UBTECH’s models, Ego (first-person-view) data accounts for around 20%, and simulation data makes up the remaining 10%.
Behind this data structure is UBTECH’s judgment about the evolution of embodied intelligence:
World models can improve data utilization efficiency, and simulation can supplement extreme scenarios. But the friction, load deformation, physical contact, and random disturbances that robots truly encounter can only be learned deeply and thoroughly through real-world interaction.
To turn this real-world feedback into a closed loop, UBTECH has built a complete self-developed data loop:
After robots enter production environments, they continuously generate task data. The data is uploaded and parsed, cleaned and labeled, used for model training, and then deployed and validated before feeding back into the next round of data collection.
During this process, every task completed by a robot is not merely a delivery—it also adds new experience to the embodied brain.
UBTECH can currently generate thousands of hours of raw data per day. After cleaning, approximately 1.5–2 hours of usable data remain for every eight hours of recorded data.
This data does not serve only UBTECH’s own models. Leading embodied-model customers have already purchased UBTECH’s real-robot data.
Financial reports show that UBTECH’s annual revenue exceeded RMB 2 billion in 2025, while total humanoid robot sales reached 13,838 units.
Among them, full-size embodied-intelligence humanoid robots generated RMB 820 million in revenue, with sales of 1,079 units. More than 80% of these robots were deployed in demanding industrial scenarios, including automotive manufacturing, smart logistics, 3C electronics, semiconductors, aviation manufacturing, and industrial data collection.
The numbers themselves are not the most important thing.
What really matters is whether these robots can enter “high-value production sites.”
The more complex the problems robots encounter there, the richer the data they generate, the faster the models can iterate, and the stronger their capabilities will be at the next delivery.
That is why the core moat in embodied intelligence has never been simple shipment volume, but rather whether a company can establish a virtuous cycle of “robot deployment → real-world data → model optimization → capability improvement → deployment in more scenarios.”
And expanding this cycle sustainably requires more than a single robotics company.
Embodied intelligence will ultimately be a systems-engineering competition encompassing product understanding, data loops, software-hardware coordination, and scaled manufacturing.
According to Jiao Jichao, UBTECH is expanding its “circle of friends” across chips, core components, complete-machine manufacturing, and scenario delivery.
At the chip level, UBTECH established “Xixuan Chuangzhi” with Montage Technology, focusing on the development and mass production of edge-side chips for embodied intelligence. Tape-out is planned for 2027, with mass production scheduled for 2028.
At the component level, it acquired A-share-listed Fenglong Co., Ltd. and is working with the company to advance the R&D and manufacturing of core components such as dexterous hands and servo drives.
△ Dexterous hand
△ Servo
On the manufacturing side, UBTECH is working with Siemens to build a super-intelligent humanoid robot factory with an annual capacity of 10,000 units, which officially began production in August this year.
In addition, on the application side, UBTECH already works with smart-manufacturing partners including Hitachi China, Foxconn, Sany Renewable Energy, Fundamental Semiconductor, and Zhucheng Technology; automotive-industry partners including BYD, Geely, Dongfeng Liuzhou Motor, Honda Trading, FAWAY, and Yongmaotai; and smart-logistics partners such as Dema Technology.
In the first half of this year, UBTECH also reached agreements with domestic and international application integrators including Shenhao Technology, Terra Robotics, Deshan JMR, Yongda Robotics, Boshi Group, and MedSci to continue promoting the deployment of humanoid robots across more industries.
Together, these partners form a complete chain:
Upstream suppliers provide more stable and lower-cost foundational robot capabilities; midstream players improve robots’ ability to be delivered at scale; and downstream real production sites continuously generate new task requirements and interaction data.
Ultimately, this data returns to the embodied brain, driving model iteration and, in turn, improving the robot’s ability to enter more scenarios.
One more thing
The humanoid robotics industry is no longer short of products that can walk, jump, or complete a single grasping action.
The gap ahead will increasingly center on a more practical question: whose robots can be retained by customers and continue working the next day.
Once robots enter real-world scenarios, they are no longer facing one-off demonstration tasks. Instead, they must perform repetitive operations for several hours a day, hundreds or even tens of thousands of times.
This tests not an isolated capability but an entire system: hardware reliability, model generalization, data-loop efficiency, and the manufacturer’s ability to understand real business operations.
From this perspective, customer orders represent more than commercial revenue. They also serve as the most direct vote of confidence in a technical path.
Customers must be willing to deploy robots before the robots can obtain real-world feedback. That feedback drives model iteration, and improved models help robots enter more scenarios.
Here, commercialization and technological evolution begin to follow the same curve.
So, looking back, the customer scenarios that UBTECH recreated at WRC…