BEIJING, Nov 2: China’s robotics industry is entering a new phase of development as leading manufacturers and technology startups shift their attention from building advanced humanoid hardware to creating artificial intelligence systems capable of making robots more intelligent, adaptable and useful in real world environments.
After gaining international attention for machines capable of running, dancing, jumping and performing complicated physical movements, Chinese robotics companies are now concentrating on the technology that could allow humanoids to perform practical tasks in factories, warehouses, hotels and other workplaces.
The emerging field, known as embodied artificial intelligence, focuses on developing software that enables machines to understand their surroundings, process information, learn from experience and respond to situations without requiring separate programming for every activity.
The shift represents a significant change in the industry’s priorities. While improvements in mechanical engineering have made humanoid robots increasingly impressive, their ability to independently understand unfamiliar situations and complete complicated assignments remains limited.
According to research firm ITjuzi, Chinese companies operating in the embodied AI sector raised 93.5 billion yuan, equivalent to approximately USD 13 billion, through 322 investment deals during the first half of 2026. The amount was five times higher than the funding recorded during the corresponding period last year.
The investment surge indicates growing confidence among investors that the next major opportunity in robotics will come from intelligent software rather than machines alone.
As competition increases and manufacturing costs decline, companies are increasingly looking towards advanced AI systems as a source of long-term commercial advantage. Industry executives believe that the ability to develop machines capable of learning different activities could determine which manufacturers eventually dominate the international robotics market.
Dermot McGrath, founder of Shanghai-based technology consultancy ZenGen Labs, said falling hardware prices were encouraging companies to focus on developing the intelligence behind robotic systems.
The competition is also becoming an important part of the wider technological rivalry between China and the United States. American companies have established strong positions in advanced artificial intelligence research and computing infrastructure, while Chinese manufacturers are seeking to use their industrial ecosystem and access to manufacturing environments to accelerate development.
Chinese startups such as X Square, Galaxea and Spirit AI are working on systems designed to help robots acquire new abilities without undergoing extensive retraining whenever they encounter a different assignment.
Spirit AI co-founder Gao Yang has described intelligence as the weakest component of the complete robotics system, highlighting the industry’s challenge of transforming impressive demonstrations into dependable commercial applications.
American companies, including Physical Intelligence, Skild AI and Generalist, are also developing AI models that can help machines learn multiple skills. Generalist has said its GEN-1.5 model completed approximately 60 per cent of simple tasks after watching demonstrations lasting between three and 12 seconds.
Meanwhile, research published by China’s X Square indicates that robots can acquire certain object-handling abilities by observing a single human demonstration, without requiring the model to be retrained for each individual task.
The developments suggest that robotics companies are moving towards systems that can transfer previously learned skills to unfamiliar activities. Such capabilities could reduce the time and expense involved in programming machines for different industrial requirements.
However, developing intelligence for physical machines presents challenges that differ considerably from those associated with conventional chatbots.
Language models can learn from enormous collections of written material available online. Robots, by comparison, require information about physical interactions, including how objects behave when lifted, how different materials respond to pressure and how machines should adjust their movements when an unexpected obstacle appears.
To address this problem, companies are collecting information through human demonstrations, wearable sensors, recorded workplace activities and computer-generated simulations.
Spirit AI has employed approximately 1,000 people to record everyday and industrial activities that can be used to train robotic systems, according to Gao Yang.
Another company, ACE Robotics, uses wearable equipment to capture workers’ movements and grip forces in commercial environments such as shops and hotels. Such information can help AI systems understand how people interact with objects and perform routine activities.
Ant Group’s robotics division, Robbyant, has also been developing its own training approach. The company said it used 160,000 hours of material, including 50,000 hours of human-guided robotic movements, to train an AI model.
Its efforts also involve testing software across 20 different robot configurations from 17 manufacturers, with the objective of developing systems that can operate across different machines rather than remaining restricted to one particular hardware platform.
DexForce, meanwhile, is using computer-generated environments to train robots. The company has reported that certain industrial assignments can be prepared for deployment within six to eight hours, with success rates exceeding 99.9 per cent in specified precision operations.
Despite these advances, experts warn that simulated environments cannot completely reproduce the complexity of physical workplaces. Real-world conditions frequently involve unexpected movements, changing surfaces, irregular objects and situations that are difficult to anticipate during software development.
Data quality is therefore emerging as an important consideration. Collecting large quantities of information will not necessarily produce better robotic intelligence if the training material fails to represent the conditions machines encounter during actual operations.
Chinese manufacturers are also facing difficulties linked to access to advanced semiconductor technology. US export restrictions have limited the availability of some high-performance Nvidia chips, which are important for developing and operating sophisticated AI systems.
The shortage could affect the speed at which Chinese companies train increasingly powerful models. Domestic semiconductor manufacturers, including Huawei, are developing alternative technologies, while robotics developers are exploring more efficient software, synthetic data and simulation-based training to reduce computing requirements.
Funding conditions also remain a concern, particularly as several startups compete in an increasingly crowded market. Questions surrounding valuations, commercial demand and the long-term profitability of humanoid robots could influence investment decisions.
Nevertheless, some companies have already begun testing their technology in practical settings.
JoyIn Technology has been experimenting with robots capable of responding to unfamiliar situations. In one example described by its development team, a robot initially failed repeatedly when instructed to place headphones inside a cupboard that was beyond its reach.
After being shown videos demonstrating how humans approached similar situations, the machine pushed a box towards the cupboard and attempted to use it as a step.
The experiment illustrated how learning from demonstrations could help robots identify alternative ways of completing tasks instead of repeatedly following unsuccessful instructions.
Industrial applications are also beginning to emerge. Battery manufacturer CATL has been using Spirit AI robots for connecting flexible wires during battery-pack testing at its Zhongzhou plant. The company has reported an insertion success rate exceeding 99 per cent.
Robbyant has separately said its robots are handling orders at two pharmacies in Shanghai, although human employees continue checking medicines before dispatch.
Such applications demonstrate that humanoid and other intelligent robotic systems are gradually moving beyond exhibitions and controlled demonstrations. However, their performance in specialised environments does not necessarily mean they are ready to undertake a wide range of unfamiliar assignments independently.
Unitree founder Wang Xingxing has suggested that future developments in large AI models could provide robotics with a breakthrough comparable to the arrival of ChatGPT in the generative AI industry.
He said the threshold for such a development would involve robots completing around 80 per cent of tasks across 80 per cent of unfamiliar environments when given spoken instructions.
Industry observers, however, caution that successful demonstrations should not be confused with commercial reliability. Businesses require machines that can consistently perform assigned activities, operate safely and handle unexpected situations without constant human intervention.
The coming years will therefore be important for determining whether the enormous investment in embodied AI can produce commercially viable systems.
For China, the challenge is no longer limited to manufacturing robots capable of impressive physical movements. The bigger objective is to develop intelligent machines that can understand instructions, learn new skills and operate reliably alongside human workers.
If companies succeed, the technology could eventually influence manufacturing, logistics, healthcare support and other industries. For now, the race to build the robotic brain has become one of the most important areas of competition in the global artificial intelligence sector.