Physical AI Enters the Scene! 2026 Industrial Expo: New Stories and Old Accounts in Industrial Robotics

“

It's about changing the way of work, but the bottleneck is the return on investment.

The 26th China International Industrial Expo opened on October 12th at the National Exhibition and Convention Center in Shanghai. The robot exhibition still occupies two entire halls, 7.1H and 8.1H, covering 56,000 square meters with over 400 exhibitors. Just walking through the two halls takes a whole afternoon.

Hall 8.1H is the international hall. The four major manufacturers—FANUC, ABB, KUKA and Yaskawa Electric—each occupy an area, with exhibition stands at the thousand-square-metre scale spread across different parts of the hall. Hall 7.1H is the Chinese hall, featuring Estun, Siasun, JAKA, AUBO and Rokae, alongside a large section of domestic component suppliers.

In previous years, walking through the middle corridor was from domestic to international, with different stories told on both sides. This year is different. I walked through both halls and noticed a detail: the word that appears most frequently in the promotional slogans of the four major families and domestic manufacturers is the same—

Physical AI

This term, which is more engineering-oriented, basically means letting robots sense, judge, and act on their own in the physical world, rather than repeating the same pre-programmed trajectory over and over. This term has been repeatedly mentioned in the context of embodied intelligence and humanoid robots over the past year, and this year it appeared for the first time in a central position on the industrial exhibition booths.

After touring with this term, I talked face-to-face with the exhibitors on-site, and we summarized

the four most notable changes, as well as three unanswered questions.

01.

Physical AI is changing the way of work

"In the past, we compared whether robots could repeat actions without deviation. Now, we compare whether they can adapt." An engineer with over ten years of experience in car welding lines summarized his feelings this year with this sentence. This can basically be taken as an explanation for this year's robot exhibition at the China International Industrial Expo. The standard for evaluating robots has changed.

The evaluation standard over the past thirty years was three numbers: repeatability accuracy, payload, and reach. It tested whether a robot could perform the same action ten thousand times without deviation.

This year, the exhibition booths repeatedly demonstrated two other questions: how long does it take to switch setups, and can it still grasp if the incoming material changes. It tested whether the robot had encountered new situations and whether it could make its own decisions.

The new products from the four major families this year are highly consistent, all based on this background.

Yaskawa Electric launched its new autonomous robot series, MOTOMAN NEXT, with the official statement being "Operation Power × Judgment Power = Robot × AI." The approach is to standardize services with vision, force, AI, and path planning, allowing the robot to complete the three steps of "understanding the situation—judgment—action" on its own.

FANUC placed "Physical AI" at the forefront of its booth, with the foundation being ROS 2, Python, and Stream Motion real-time motion control. On-site demonstrations included multimodal visual perception, autonomous obstacle avoidance, and real-time trajectory control.

ABB is talking about "industrial-grade physical AI" from concept to implementation, demonstrating through a manual coffee station, where the GoFa robot observes human movements and can replicate the trajectory and pouring rhythm of the kettle. The base is RobotStudio HyperReality and OmniCore EyeMotion, one generating synthetic data for virtual validation, and the other reducing the visual deployment cycle from months to days.

KUKA launched the "KAI" intelligent platform, with the robot body equipped with the iiQKA operating system, allowing AI decision-making to run directly on the device without sending data back to the cloud for instructions.

Domestic manufacturers answered the same question.

ESTUN showcased the Lingzhi dual-arm force-controlled robot, using single-controller multi-machine coordination, multimodal fusion, and millisecond-level perception contact to solve issues such as dual-arm coordination debugging, multi-machine interference avoidance, and difficulty in ensuring force control consistency, targeting production lines with random incoming materials and frequent setup changes.

Aritel's booth, themed "One Brain, Many Shapes, Smart Chain Without Boundaries," presented the strategic results of embodied intelligence moving from technical validation to large-scale industrial implementation. The highlight workstation used five humanoid dual-arms for optical module insertion, inspection, and transportation processes, with multiple robots scheduled based on a unified embodied large model, capable of precise assembly and adaptive grasping at sub-millimeter precision. The optical module insertion space is narrow and the precision requirements are high, relying on the robot's own judgment on how to insert, rather than following pre-set points.

The booths of six companies are located in different corners, but they all share a common point:

The change is concentrated in the "perception + judgment" layer.

This layer. The hardware body layer this year had almost no surprises; everyone was competing on the ability to work under uncertain conditions. Teaching playback has not disappeared, it has just moved from being the only way of operation to one of the options.

02.

The control cabinet is becoming an open platform

"Customers used to buy a determined rhythm, now they buy a system that needs to be raised together."

The person who said this is a project manager in integration. He does not oppose AI, but he opposes equipping every process with a brain.

But before explaining this debate clearly, we need to look at the most concrete outcome of this round of changes, the control cabinet.

Previously, the control cabinet was the territory of the manufacturers. Process knowledge was locked within the manufacturer's proprietary language and development environment, and customers wanting to modify a process had to turn to the manufacturer or its authorized integrator. This year, the four major players have all opened up their cabinets.

Yaskawa's YNX1000 supports developing robot applications using general-purpose languages such as C++ and Python, eliminating the need to learn the manufacturer's proprietary language. Four types of services—vision, AI, force sensing, and path planning—are open for external use, accompanied by SDKs, development kits, and simulators, allowing development and verification without the need for actual hardware. It is also integrated with NVIDIA Isaac Sim for digital twin applications.

FANUC's open platform also uses ROS 2 and Python as entry points, ABB brings the E-Device programming tool, and KUKA releases the Chinese version of Visual Components, all aiming to do the same thing: making process knowledge and application development no longer locked within the manufacturer's control.

For customers, this directly corresponds to the most expensive hidden cost in industrial robots—

Integration and debugging

. The real selling point of an open platform is not technological advancement, but transforming the work that was previously only entrusted to the manufacturer into something that customers and third parties can also handle.

However, openness is just the prerequisite; whether or not to install a "brain" is another matter. The "little brain" and "big brain" are treated differently on the exhibition booth. The "little brain" is a decades-old friend, with the servo drive in the control cabinet doing its job; the "big brain" is a new arrival this year, bringing computing power, models, and a string of unresolved costs.

One group sees the "big brain" as a must-have. An engineer working on motion control told me: "The work on the automotive line is standard, with tolerances, fixtures, and cycle times locked in place. Pre-programmed trajectories can run for years without error. However, in the workshop, there are always some processes that don't follow the standard, with each batch of incoming materials having differences, fixtures wearing out, and room temperature varying between summer and winter. These tasks the "little brain" can't handle, as it only recognizes pre-programmed trajectories. The "big brain" does is absorbing these variations and recalculating on the spot."

His judgment is that

the higher the precision requirements and the more chaotic the incoming materials, the more unavoidable the "big brain" becomes

. There's another layer he considers more important. Industrial robots have been doing work for thirty years, with scenes basically locked in automotive and electronics. Other industries are not without demand, but the work is too scattered and too expensive to hire an integrator. The "big brain" truly opens up this market.

Another group doesn't accept this argument. "Count the number of tasks in the workshop, and most are still repetitive work. Pre-programmed trajectories run fast and stable, and if they break, they can be repaired quickly. Why give these tasks a "big brain"? Computing power is money, data collection is money, and the most expensive is debugging. If the model has a problem, the engineer has to be on-site."

A project manager in integration

calls this "overkill" brain, not that the brain is bad, but that most tasks still don't need it.

The two groups are not arguing about whether the brain is good or bad, but rather about where to start and how deep to go. The clue lies on the customer side, with more people asking about AI capabilities, while those asking about the payback period remain the same.

03.

Cost: the upward additions and downward pressures are currently offsetting each other

"A machine that can't recoup its costs will be dismantled on the day of its replacement," said an engineer in charge of production line transformation at a manufacturing company. They have an internal payback period red line, and if it's exceeded, they can't even get into the pilot phase.

Physical AI is not a free lunch. Computing power needs to be added, force sensing sensors need to be added, and vision needs to be added, with the cost of hardware materials going up.

But another clear fact at this exhibition is that the upstream is pushing harder downward, with the two lines forming an offset.

The composition of the robot exhibition is the most straightforward evidence. There are 108 whole robot manufacturers and 169 core component manufacturers, with the upstream being denser than the downstream. Companies like Green Harmonic, Lai Fu Harmonic, Huan Dong Technology, Tongchuan, and Weiling Robot Parts, which produce reducers and joint modules, are all gathered in two exhibition halls.

In addition, the density of the supply chain is not only reflected in horizontal clustering.

Huayan

Robot

is based on over 20 years of core technology accumulation, achieving self-reliance and self-sufficiency in core motion components such as motors, servo drives, and joint modules, while also possessing the ability to independently sell externally.

The result of the offset is reflected in two sets of numbers.

First is the market share

According to MIR Dornier, the domestic industrial robot market share of Chinese brands in the first quarter of 2026 reached 55%.

Second is the gross margin, ESTUN's comprehensive gross margin for the first half of the year was 31.81%, an increase of 4.17 percentage points year-over-year, and the company attributed this to product structure upgrades and raw material domestic substitution, among other cost-saving and efficiency-enhancing measures in its semi-annual report.

However, market share and gross margin are matters for the leading companies; the industry as a whole has not yet reached the point where it is profitable.

Similarly, in the first half of the year, XinSong's revenue was 1,451,000,000 yuan, a decrease of 12.57% year-over-year, with a net loss of 189,000,000 yuan for the parent company, experiencing both scale contraction and profit pressure. YueXiang's revenue was 316,000,000 yuan, doubling year-over-year, but with simultaneous increases in expenses. ESTUN chose the third path, actively lowering its gross margin and homogenized orders, achieving revenue of 1,641,000,000 yuan in the domestic market, a decrease of 8.82% year-over-year, resulting in a positive non-GAAP net profit and positive operating cash flow.

Putting these three together illustrates the same thing, In this cycle, selling the product is not difficult, but getting the accounts right is becoming increasingly difficult.

Asked about this expense at the exhibition booth, the person in charge of equipment gave a more restricted explanation than the manufacturer. It was still the engineer in charge of production line renovation who told us

"Robots are used to replace production capacity; production lines need to be upgraded every three to five years. 36 months is the hard limit; if you push it down to under 24 months, procurement will be more willing to loosen up."

Those doing integration are accounting for another set of numbers. A project manager said: "Customers may ask how much a robot costs, but what they truly pay for is three parts: hardware, debugging, and maintenance. Hardware is paid for once, debugging is charged per person-day, and the most expensive is on-site support; models are not finalized upon delivery, and when materials change, data needs to be re-collected and re-adapted." He has seen too many projects, where the price was quoted for a single piece of equipment during procurement, but the requirements were raised for an entire system during acceptance, and the difference in cost was left unaccounted for.

There is also a colder voice. An engineer who does intelligent transformation for factories said: "Currently, there is no production line that runs at 100% capacity. If precision is insufficient, it is compensated with force control; if stability is insufficient, a remote takeover position is left; as for success rate, it is negotiated with the customer on fault tolerance. Every running robot is a compromise, and balancing the accounts is more important than maximizing the parameters."

Putting these accounts together, Physical AI adds not only cost to the material list of industrial robots, but also variables to the customer's return-on-investment table.

Every penny pressed down from the upstream is replacing downstream with a bit more room for trial and error. However, as long as the return-on-investment cycle remains the only hard metric, leading parameters on the spec sheet can't be temporarily converted into premium pricing.

04.

Going Global: Physical AI is a new selling point, as well as a new barrier to entry

"The first round of questions were about parameters, the second round about what AI can do, and the third round was about price," said a company official responsible for overseas business, describing the change in the order of questions from European customers this year.

Chinese companies have been installing robotic arms in factories in Germany, Japan and the United States for around ten years. This year's new variable is physical AI, which is changing both ends of their overseas expansion.

On the sales side, it is a new opening line; customers have raised AI capabilities to the first round. On the entry side, it is a new exam question; the more a robot can judge on its own, the more things it needs to be proven.

Combining both ends, going global has transformed from a short sprint of price competition into a long-term battle of service and compliance.

The basic export market is still expanding. China's collaborative robots' exports, for the first time last year, exceeded 10,000 units, with customs data showing 10,872 units, of which Germany purchased 2,360 units. In the first half of this year, exports of industrial robots reached 6,290,000,000 yuan, a year-over-year increase of 18.6%, shipped to 141 countries and regions; the average price per unit of robots exported through the Shanghai port rose to 86,000 yuan, with an 84% increase in exports to Germany and a 36.5% increase in exports to Japan.

Taking Physical AI capabilities abroad, several companies have clearly accelerated their overseas layouts this year, moving from simply selling equipment to building bases, seeking funding, and expanding product lines.

YueXiang's total revenue for the first half of 2026 is approximately 316,000,000 yuan, doubling (+106.6%), with overseas revenue of approximately 153,000,000 yuan, accounting for 48.36%. CEO Liu Peichao revealed that the company currently derives about 60% of its revenue from overseas, with products already entering over 100 countries and regions. Huayan listed on the Hong Kong Stock Exchange on March 30, with a 5,059 times oversubscription in the public offering, with about 20% of the raised funds clearly allocated to overseas business. Ailite has laid out products according to three major product matrices: collaborative, compound, and humanoid, with products sold in over 50 countries. ESTUN's overseas revenue for the first half of the year was 936,000,000 yuan, a year-over-year increase of 25.06%, with the proportion of total revenue rising to 36.33%, and the warehouse manufacturing base in Poland was put into operation in the first half of the year.

But after robots have been sold to 141 countries, the next step for exports is not orders, but service. Where spare parts are stored, how quickly engineers can arrive, and who handles local certification will all directly affect repurchase rates.

The threshold is also rising. The new EU Machinery Regulation will be fully enforced in January 2027, and industry estimates suggest compliance costs will rise; the safety requirements for collaborative robots have been fully incorporated into the new ISO 10218:2025 standard, with new regulations on force limits and speed monitoring further increasing technical entry barriers.

Physical AI makes robots smarter at the same time, it also needs to prove that its judgment is reliable, that responsibility is clear in case of incidents, and that data can be stored and shared.

An opening statement can win attention, but exam questions decide whether you can enter. Safety responsibility and data compliance must follow the robots on their journey abroad.

05.

Three questions that still don't have answers

Beyond the trends, the truly valuable information from this exhibition lies in the questions. The more lively the explanations on the exhibition booths, the bigger the question marks often are.

The first is the data gate, which hasn't been passed yet.

To make models smarter, real machine data is needed, but collecting real machine data is expensive and not transferable across different robot bodies. On-site, some companies have even turned this into a product. Meifeng Technology first showcased the MEgo series of no-body data collection terminals and the physical AI data crowdsourcing platform “Meifeng Pai,” attempting to temporarily separate data collection from the robot body through governance and mapping to allow data to enter the model; Digu Robot showcased the RoboGo full-chain platform covering data collection, training, verification, and deployment.

These solutions represent the direction of solving the problem, but data collection still relies on specialized hardware and manual organization, making large-scale implementation difficult; and whether the data collected can effectively migrate across different robot bodies remains without public quantitative verification. Cross-body migration and few-shot learning have already been applicable in structured scenarios, but non-structured conditions still rely on manual intervention.

The second is the question of “how long to break even,” which is still the most frequently asked question in front of the booth.

Manufacturers talk about computing power, models, and degrees of freedom, while customers calculate cycle times, yield rates, and downtime. Leading parameters on the parameter sheet can't temporarily be converted into premium pricing.

The third is the gap between demonstrators and industrialized products.

Robots that appear stable on the exhibition booth may suffer from culture shock when entering dusty, high/low temperature, and high-vibration workshops. The lifespan of components, temperature drift, and anti-interference capabilities are repeatedly mentioned as shortcomings on-site; in terms of heavy-load RV reducers, high-precision force sensors, and encoders, the verification cycle for high-end scenarios is still very long. In high-end applications such as automotive production lines, a significant proportion of the large-load machine types still rely on imported reducers.

Data, break-even, and reliability—these three things aren't something that can be crossed over with a single technological upgrade; they only recognize time and hours. The industry is still in the process of turning technology into products.

Putting several changes together, the transformation of industrial robots by physical AI is currently mainly happening at two levels:

Operation methods, from teaching and repetition to autonomous judgment; Development methods, from closed control cabinets to open platforms. Above that, the real deciding factor for whether this business will succeed is still the old three numbers— Cost, delivery, and repeat purchase.

The corridor between 7.1H Hall and 8.1H Hall, which used to be a path from domestic to international, is now both sides talking about the same set of terms. The only difference is: who has already run for how many hours on the customer's production line.

END

Industrial robot companies

Eston Automation

Efte Robot

Fao Robot

Yuexiang Robot

Jieka Robot

Songling Robot

Luoshi Robotics

A Tongmu Robot

JiZhiJia

Hikvision Robot

Yifei Technology

Ailite Robot

Service and special robot companies

Yijiahe

Jingpin Tieshang

Qiteng Robotics

Shihe Robot

Pudu Robot

Shiluo Robot

Kuma Technology MAMMOTION

Humanoid robot companies

UBTech Robotics

Unitree

Yunchedu

Xingdongjinyuan

Weijing Robot

Zhujidongli

Leju Robot

Daxiang Robot

Magic Atom

Zhongqing Robot

Pasi Ni Perception

Cyborg Robot

Digital Huaxia

Fourier Intelligent

Tianlian Robot

Kepler Humanoid Robot

Lingbao CASBOT

Qingbao Robot

Zhejiang Humanoid Robot Innovation Center

Dongyi Technology

Zhishi Technology

PNDbotics

Zhuoyide Robot

Qinglang Intelligent

Galileo GALILEO

Songyan Power

Tianji Intelligent

Woaan Robot

LiGong Huahui

Jiace Jinhua

Embodied intelligence companies

Cross Dimension Intelligent

Galaxy General

Qianxun Intelligent

Lingxin Qiaoshou

Ruierman Intelligent

Microintelligence

Tuisheng Technology

Zhongke Silicon Ji

Shutu Technology

Lingqiao Intelligent

Xingchen Intelligent

Qiongche Intelligent

Fangzhou Infinite

| iFLYTEK |

Beijing Humanoid Robot Innovation Center

National and Local Co-built Humanoid Robot Innovation Center

Daimeng Robot

Shibitu Robot

Xinghaitu

Yuequan Bionics

Zeroth Power Robot

Zhongke ShenGu

Zhi Fangping

Daka Robot

Haocun Technology

Ju Shi Intelligent

Xynova Xiner Future

Feixi Technology

Weilu Power

Boden Intelligent

Qianjue Technology

Lingsheng Technology

Ji Cui Intelligent Manufacturing

Xinbeit Technology

Chen Hunxian Technology

Dexmal Original Power

Youliqi

Self-variable

Ruiyan Zhikong Lingqiao Hand

Qiwu Technology

RoboScience Machine Science

Zhongke Fifth Dynasty

Critical Point

Danghong Technology

Qiaojie Shu Wu

Vbot Veta Power

Tashan Technology

Ju Brain Pan Shi

Youai Zhijhe Robot

Zhixing Jian

Aimiao Robot

Lu Ming Robot

Quantum Intelligent

Ruoyu Technology

Tiehu Robot

Zhi Wuji

Medical robot companies

Yuanhua Intelligent

Tianzhihang

Sizhe Rui Intelligent Medical

Jingfeng Medical

Tuo Dao Medical

Zhenyida

Shuru® Robot

Luosenbot

Shuimu Dongfang

｜

Kangnuo Sistem

Disight Medical

Upstream industry chain companies

Lvde Harmonic

Yinshi Robotics

Kunwei Technology

Muta Intelligent

Qingpou Vision

Benmo Technology

Lan Dian Touch Control

| Xinjingcheng Sensor

BrainCo Strong Brain Technology

Yuli Instrument

Jiya Jingji

Sila Technology

Shenyuan Sheng

Feipu Navigation Technology

Inks

Juxi Intelligent Drive

Lingyun Guang Yuanke Vision

Xuanji Power

Yiyou Technology

| Ruiyuan Precision |

Lingzhu Era

HIT Huawei Ke

Xinghui Sensor

Lingdi Technology

Quan Zhi Bo

CubeMars Robot Power

Wanglong Robot Elevator

Hangkai Microelectronics

Wutong Sensing and Control