Physical AI in factories: what China teaches manufacturers in India and Singapore
In March 2026, a two-armed robot on wheels passed its acceptance test at CATL's battery plant in Ningde, in south-east China. Since then it has moved materials on the module and pack lines around the clock, picking up work that people used to do by hand. It was not programmed move by move. It learned.
Stories like this one travel fast, and they reach owners of manufacturing companies in Pune, Coimbatore and Jurong in the form of one question: is this coming to my shop floor, and do I need to act now?
To answer it, we went to where this industry is moving fastest. We read the data from the International Federation of Robotics, research firms, government agencies in China, India and Singapore, and the factories that have published their results. Then we read it all through the eyes of an owner running a company with 50 to 300 people. We say where the data is solid, where it is thin and where the conclusion is ours.
At the end you will find a tool, the task map, to work out which jobs in your company are ready for a robot today and which are worth watching for the next few years.
What is physical AI, and why is everyone talking about it now?
Physical AI is artificial intelligence that controls a body: a robot that learns a task from examples and data, then adapts when the part or its position changes. Researchers also call it embodied AI.
The technical core is a family of models called vision-language-action models. They take in camera images and an instruction in plain words, and return the movements of arms and hands. A traditional industrial robot repeats a path someone programmed. A robot running one of these models can cope when a box arrives at a slight angle, or a fabric panel folds differently.
What changed in the last two years is data. Robots learn from demonstrations, and demonstrations take hours of real work to record. China has opened dedicated training centres where robots practise factory and warehouse tasks all day, such as a facility launched in Cixi, Zhejiang, in May 2026 (Xinhua, July 2026). More data, more versatile robots: the same dynamic that made language models useful, applied to hands.
A note on the evidence. Most performance figures for learning robots come from tests run by the companies that build them. They show the direction clearly. How a robot performs on your line can only be measured on your line.
How many robots are really at work, in China and the world?
In 2025 the world installed just over 600,000 industrial robots, 11% more than the year before, and the number at work in factories reached a record 5 million (IFR, World Robotics 2026). China alone installed 354,000, or 59% of the total, and Chinese brands supplied 55% of them. The United States was second with about 38,500 units, then Japan with 36,219 and South Korea with 30,000. India came sixth, with almost 10,500.
| Key figure (IFR, World Robotics 2026) | 2025 |
|---|---|
| Industrial robots installed worldwide | 600,000+ (up 11%) |
| Installed in China | 354,000 (59% of the world) |
| Installed in India | almost 10,500 (up 15%, sixth in the world) |
| Robots at work in factories worldwide | 5 million (up 9%) |
| Robot density, Singapore (robots per 10,000 manufacturing employees) | 873 (second in the world) |
| Robot density, China | 193 (19th in the world) |
| Robot density, world average | 152 |
Robot density, the number of robots per 10,000 manufacturing employees, completes the picture. In its November 2026 release, the IFR puts South Korea first at 1,066, Singapore second at 873, Germany third at 469 and Japan at 452. China is 19th at 193, based on new employment data from its national economic census. Our reading: China installs more robots than anyone because it has an enormous workforce to support, and it still has a long road ahead.
Humanoid robots are a much smaller world. Over 22,000 shipped worldwide in the first half of 2026, almost four times as many as a year earlier, and Counterpoint Research expects more than 50,000 for the full year (Counterpoint Research, August 2026). The top five makers are all Chinese: AgiBot with about 9,700 units, Unitree with over 7,000, then Galbot, UBTech and Leju.
For a manufacturer, the most useful number is where those robots go. More than 60% end up in entertainment, research and data production, about 19% in services and guided tours, 13% in manufacturing and 5% in warehousing and logistics (Counterpoint Research, August 2026). Research firms disagree by a few thousand units on the totals. The European think tank MERICS counts 12,800 humanoids produced in China in 2025, about 90% of the world total, against 556,000 industrial robots made in the same year (MERICS, April 2026).
What are learning robots actually doing in Chinese factories?
They do single, clearly bounded tasks, with results the companies themselves measure and publish.
Battery plants. At CATL's Ningde base, Galbot's S1, a wheeled robot with two arms and a 50 kg payload, passed acceptance in March 2026 and has since handled materials on module and pack lines around the clock, according to CATL (reported by Gasgoo, June 2026). The two companies have signed an agreement to scale it up and bring it into CATL's service network. CATL had earlier put a different robot, Xiao Mo from Spirit AI, to work on end-of-line battery pack testing.
Car plants. In Xiaomi's electric vehicle factory, a humanoid fitting self-tapping nuts reached a 90.2% success rate over three hours of continuous work, at 76 seconds per cycle, which matched the station's tightest cycle time (Xiaomi, March 2026). By July, after four months of iteration, Xiaomi reported 98% at the same station, one point below its human workers. The robot is still a prototype in internal validation.
Textiles and components. Aitu, a sewing technology company in Zhejiang, uses robots to separate fabric pieces and reports a 97% success rate for one product and fabric type, with payback in about 18 months (Xinhua, July 2026). On a seat-belt component line, PIABOT's G2 robot taps aluminium parts in as little as 12.9 seconds per cycle, against about 18 seconds for a worker. UBTech, which works mainly with car and electronics makers, raised its 2026 forecast to 5,000 Walker S humanoids (Counterpoint Research, August 2026).
The makers themselves are clear about the limits. Lou Lingtong, general manager of Aitu, told Xinhua that humanoid robots "are still at the learning stage". He expects 2026 to 2028 to be spent validating single processes, with coordinated use across several tasks from 2028 to 2030. The obstacles they list: soft materials, reflections and dust that confuse cameras, wear and heat in continuous work, and the lack of common interfaces and safety standards. MERICS adds that most humanoids in China still run "small-scale, single-point trials", and quotes UBTech's founder saying they are only half as efficient as people.
A note on the evidence. Success rates, cycle times and payback periods above are company figures, reported by news agencies and trade press. None has been independently verified.
Our reading: the cases that work today share three traits. Parts that are similar to one another, a task that repeats for hours, and a person nearby ready to step in. Those are the same conditions in which a company in India or Singapore can start now, with simpler robots.
Who pays for China's push, and what does it mean for buyers?
The state steers and pays. The IFR notes that China's 15th Five-Year Plan (2026 to 2030) points its AI research at physical applications, with robots as a main driver of growth (IFR, September 2026). MERICS counts a national venture fund and three regional funds worth CNY 1 trillion, about €120 billion, over 20 years, and local subsidies of up to 30% of project costs for automation (MERICS, April 2026). In June 2026, China's industry ministry and its state assets regulator launched a programme to train humanoids in real workplaces, targeting deployment at the scale of 10,000 units by the end of the year (Xinhua, July 2026).
So much push has a price. MERICS describes humanoid makers "waging a price war" and records a warning from China's national planning commission about possible overcapacity. For a buyer, the practical consequence is good news: prices are falling fast.
How much does a learning robot cost, and when does it pay back?
According to MERICS, a complete Chinese industrial humanoid costs on average CNY 300,000 to 500,000, about €36,000 to €62,500. The small Unitree G1 costs around €11,650. On top of the robot come integration, safety work and training.
MERICS also cites a simple rule from the Chinese broker Guotai Securities: a humanoid becomes commercially viable when it costs no more than two years of the wages of the worker whose task it takes on. The rule is crude, but it explains why the same robot makes sense in one market and not in another.
| Two years of one operator's wage | Approximate amount | Compared with a Chinese humanoid (€36,000 to €62,500) |
|---|---|---|
| Singapore, metalworking machine setter-operator (median gross wage S$2,718 a month, plus 17% employer CPF) | about S$76,000 | Roughly at or above the price, about S$54,000 to S$94,000 |
| India, average salaried manufacturing worker (₹18,735 a month) | about ₹4.5 lakh | A small fraction of the price, well over ₹30 lakh |
Sources: MOM Occupational Wages 2025 (June 2025, full-time residents); PLFS 2025 via NCAER; MERICS. Euro conversions at about S$1.50 per euro. This is our own illustration, not a forecast.
What this means, in our reading:
- In Singapore, the wage case for robots is already strong. Resident wages are high, foreign worker levies add to the cost, and the IFR ranks Singapore second in the world for robot density. Learning robots will compete on price within a few years.
- In India, labour cost alone will not justify a humanoid for a long time. The case rests on other things: quality and consistency that export customers demand, hard-to-staff roles, safety, and scale. For most MSMEs, the right first step is a cobot or a simpler fixed robot, not a humanoid.
Where does Singapore stand?
Singapore starts from one of the most automated factory floors in the world, and its government is now funding AI on that floor directly.
Density. Singapore has 873 robots per 10,000 manufacturing employees, second only to South Korea (IFR, November 2026). The IFR notes that a small manufacturing workforce helps push the ratio up.
AI adoption. The gap is AI, not robots. In 2025 only 18.0% of Singapore's manufacturing firms used AI, against 46.8% in information and communications, according to IMDA's Singapore Digital Economy Report 2026. We looked at what that report means for B2B owners in our guide to the SGDE 2026.
Catalytic AI Solutions. On 21 October 2026, Enterprise Singapore committed more than S$45 million over three years to help manufacturing SMEs adopt ready-made AI solutions instead of building their own (Enterprise Singapore, via TechNode, October 2026). The solutions come from A*STAR's Sectoral AI Centre of Excellence for Manufacturing and are offered through trade associations and qualified vendors. Named use cases: faster product design, automated request-for-quotation handling, machine process and quality inspection, real-time energy monitoring, production scheduling and predictive maintenance.
The Embodied AI Challenge. Launched the same week, the Singapore Embodied AI Challenge for Manufacturing is run by the National Robotics Programme, A*STAR and PINE Lab at NTU, with Micron as title sponsor and Galbot supplying the robot platform. Teams teach a robot to open, service and reassemble a semiconductor equipment chamber. Registration closes on 31 December 2026 and the finals are on 20 March 2027 (A*STAR, October 2026).
Grants for robots. Under the EDGE Grant, which replaced the Productivity Solutions Grant in September 2026, a robotic arm is a pre-scoped item: SMEs can get up to 50% of a cost capped at S$35,000, so up to S$17,500, within an annual automation and digitalisation cap of S$30,000 (Enterprise Singapore). The company needs at least 30% local shareholding, and no payment can be made before the application.
A local example. Seng Heng Engineering, a third-generation Singapore maker of fasteners for oil and gas equipment, put two cobots on CNC loading and unloading and part engraving when border closures left it short of staff. Universal Robots, the cobot maker, reports a 50% gain in line productivity within three months. It is a vendor case study from the pandemic years, but it shows the pattern: a repetitive task, a clear bottleneck and people moved to better work.
Where does India stand?
India is one of the fastest-growing robot markets in the world, from a low base, and its learning robots are still in the lab and on the trade-fair floor.
Installations. Indian factories installed almost 10,500 industrial robots in 2025, up 15%, making India the world's sixth-largest market (IFR, September 2026). Installations grew 27% a year on average between 2020 and 2025. The car industry leads: in 2024 it took about 45% of installations, and installations at parts suppliers rose 40% to 2,100 units (IFR, September 2025).
People. In ManpowerGroup's 2026 survey, 82% of Indian employers said they struggle to find the skills they need, and 79% in manufacturing (ManpowerGroup, 2026). India's problem is rarely the number of people. It is finding, training and keeping skilled ones.
Policy. India does not yet have a robotics scheme an MSME can apply to. A draft National Strategy on Robotics exists, and in February 2026 an advisory group under the Principal Scientific Adviser proposed a National Robotics Mission, shared training data, testing labs and production incentives for robot components (Dataquest and KNN, February 2026). These are proposals. What MSMEs can use today:
- Mutual Credit Guarantee Scheme (MCGS-MSME): a 60% guarantee from NCGTC on loans of up to ₹100 crore for equipment and machinery. Since March 2026, equipment must make up at least 60% of the project cost, down from 75%.
- Union Budget 2026-27: a ₹10,000 crore SME Growth Fund for equity in future "Champion MSMEs", and a scheme to revive 200 legacy industrial clusters (Government of India, February 2026).
Learning robots. Addverb, the Noida robotics company owned by Reliance, unveiled Elixis-W, a wheeled humanoid for warehouses and factories, at LogiMAT India in February 2026, with pilots planned with industrial partners. As of November 2026 we found no public, measured deployment of a learning robot on an Indian production line. That is not a reason to wait. It means Indian MSMEs have time to prepare tasks and data while prices fall.
The task map: which jobs to give a robot, and when
The task map scores each repetitive job in your company and tells you whether it is ready for a robot today or worth watching for the next few years. Take one task at a time, for example loading a press, palletising cartons or a visual check, and answer five questions with a score from 0 to 2.
Before you start. For each task, note whether cycle times, volumes and rejects are already recorded. Without those numbers no robot can be justified, and no learning robot can be trained.
Pain: how much the task costs you today
| Question | 0 points | 1 point | 2 points |
|---|---|---|---|
| 1. How often does it repeat? | Changes every day | A few hours a day | Whole shifts |
| 2. How hard or risky is it? | Light work | Tiring | Heavy, dirty or dangerous |
| 3. How hard is it to staff? | Filled quickly | Takes months | Stays open, or people leave fast |
Variety: how much the robot needs to understand
| Question | 0 points | 1 point | 2 points |
|---|---|---|---|
| 4. Parts and positions | Always the same | A few part families | Change almost every time |
| 5. Judgement | None needed | A visual check | Constant decisions |
How to read your score
- Pain 4 to 6, variety 0 to 1: ready today for a traditional industrial robot or a cobot. In Singapore, check the EDGE pre-scoped equipment list; in India, ask your bank about MCGS-MSME.
- Pain 4 to 6, variety 2 to 4: physical AI territory. Start recording data now and plan a pilot for 2027 to 2028, when learning robots are cheaper and better proven.
- Pain 0 to 3: leave it with your people and look again in a year.
The rows with high pain and no recorded data are where to start. That is where you lose hours every week without being able to measure them.
What should owners do in the next 12 months?
- Fill in the task map with your supervisors, on at least ten jobs. It takes about two hours.
- Measure the high-pain tasks: cycle times, volumes and rejects. Numbers beat impressions when you talk to a vendor, a bank or a grant officer.
- Start with the tasks that are ready today, using proven industrial robots or cobots and the support that exists now: EDGE in Singapore, MCGS-MSME in India.
- Choose who will look after the robots and train them in maintenance and set-up. Chinese researchers quoted by Xinhua expect robot maintenance, debugging and set-up to create new jobs.
- Ask every vendor how their robot will be supported locally, who integrates it and what safety standards it meets. Common interfaces and standards are still missing, so service matters more than the spec sheet.
- Tell your buyers. A factory that automates with method has a strong story for customers, especially export customers who audit their suppliers. Buyers check consistency, capacity and quality long before they call, as we explained in how manufacturing buyers choose suppliers.
Where does EGGA fit?
EGGA does not sell or install robots. We are a growth, sales and marketing service for SME owners, on one fixed monthly fee. For manufacturers, that means turning real operational strengths, including automation, into the proof buyers look for: a website and content that show capacity and quality, visibility on Google and AI search, and a steady flow of qualified enquiries.
If you want to see where your company stands today, the Growth X-Ray gives you a free, research-based diagnostic. Plans and prices for India and Singapore are on our pricing pages.
The head start belongs to the factory that knows its tasks
China chose to sprint, with public money, dedicated hubs and huge volumes of training data. The humanoids it produces today do single tasks with good results, and their own makers put whole-line use between 2028 and 2030. That is useful time.
Singapore enters this period with one of the most automated factory floors in the world and new money to put AI on it. India enters it with one of the fastest-growing robot markets in the world and a skills gap that makes automation a matter of quality, not only cost. In both markets, the owners who measure their tasks today, automate the first ones and train the people who will run the machines will be ready when learning robots become affordable. The head start belongs to the factory that knows its tasks.
Questions people also ask
When will humanoid robots reach factories in India and Singapore?
Pilots are already running in Chinese factories, and Singapore's Embodied AI Challenge puts a Galbot robot to work on semiconductor equipment in March 2027. Chinese makers expect coordinated use across several tasks between 2028 and 2030 (Xinhua). For most SMEs in India and Singapore, 2027 to 2028 is the sensible window for a first pilot on one task.
How much does a humanoid robot cost in 2026?
According to MERICS, a Chinese industrial humanoid costs on average CNY 300,000 to 500,000, about €36,000 to €62,500, while the small Unitree G1 costs around €11,650. Integration, safety work and training come on top. Prices are falling quickly because Chinese makers are competing hard on price.
What is the difference between a cobot and a robot with physical AI?
A cobot, or collaborative robot, follows programmed movements and is designed to work safely next to people. A robot with physical AI learns a task from demonstrations and data, and adapts when parts or positions change. Cobots are mature and grant-eligible today; learning robots are still in single-task trials, mostly in China.
Is there a grant for buying a robot in Singapore?
Yes. Under Enterprise Singapore's EDGE Grant, a robotic arm is a pre-scoped item: SMEs can receive up to 50% of a cost capped at S$35,000, so up to S$17,500. It counts towards a S$30,000 annual cap for automation and digitalisation. The company needs at least 30% local shareholding and must apply before paying any supplier.
Is there government support for MSMEs buying robots in India?
There is no robot-specific scheme yet: a National Robotics Mission has been proposed but not launched. MSMEs can use the Mutual Credit Guarantee Scheme, which gives a 60% guarantee on loans of up to ₹100 crore for equipment and machinery, where equipment is at least 60% of the project cost. Check the terms with your bank.
Will robots replace factory workers?
In the cases published so far, robots take over single, repetitive or heavy tasks, and people stay to run, check and maintain them. Aitu's general manager says robots cannot yet run a whole production flow without humans (Xinhua). Chinese researchers expect new jobs in robot maintenance, debugging and set-up, which is where training budgets should go.
Sources
- Five Million Robots now Operate in Factories GloballyInternational Federation of Robotics, World Robotics 2026, 24 September 2026, 2026
- China Installs 59 Percent of World's Industrial RobotsInternational Federation of Robotics, 24 September 2026, 2026
- Robot Density Highest in South Korea, Singapore and GermanyInternational Federation of Robotics, 12 November 2026, 2026
- India rises to sixth in global factory robot installationsThe Tribune, on IFR World Robotics 2025, September 2025. Secondary source, 2025
- Global Humanoid Robot Shipments Soar Nearly 300% YoY in H1 2026Counterpoint Research, 20 August 2026, 2026
- Embodied AI: China's ambitious path to transform its robotics industryWendy Chang, Rebecca Arcesati, Altynay Junusova, MERICS Report, April 2026, 2026
- China's humanoid robots face their real test on the factory floorXinhua, 24 July 2026, 2026
- Galbot's Galbot S1 begins operations at CATL's factoryGasgoo, on a CATL statement, June 2026. Secondary source, 2026
- Xiaomi updates progress on humanoid robots in auto factoryTechNode, on Xiaomi data, 15 July 2026. Secondary source, 2026
- Enterprise Singapore allocates S$45m to speed AI adoption in advanced manufacturingTechNode Global, on an Enterprise Singapore statement, 21 October 2026, 2026
- Singapore launches first national Embodied AI Challenge to accelerate real-world robotics deployment in manufacturingA*STAR, 26 October 2026, 2026
- EDGE Grant: Robotic armsEnterprise Singapore, consulted 16 November 2026, 2026
- Singapore Digital Economy Report 2026IMDA, 5 October 2026, 2026
- Occupational Wages 2025Ministry of Manpower, Singapore, Table 1, June 2025 data, 2025
- Seng Heng Engineering case storyUniversal Robots. Vendor source
- Manufacturing penalises experienced workersJayanta Talukder, NCAER, on PLFS 2025, 2 May 2026, 2026
- Global Talent Shortage Survey: India report 2026ManpowerGroup India, 2026, 2026
- Govt reviews India's robotics ecosystem, eyes national roadmap for physical AIKNN India, 4 February 2026, 2026
- India outlines strategic roadmap for self-reliant robotics and physical AIDataquest, 4 February 2026, 2026
- Union Budget 2026-27 highlightsPress Information Bureau, Government of India, February 2026, 2026
- Mutual Credit Guarantee Scheme for MSMEs (MCGS-MSME)Indian Bank, scheme page; changes per Ministry of Finance press release of 21 March 2026, 2026
- Addverb unveils Elixis-W and intralogistics solutions at LogiMAT India 2026TimesTech, February 2026. Secondary source, 2026