Trending
Chinese companies are investing heavily in artificial intelligence designed to make humanoid robots more useful in everyday and industrial settings.
After years of developing robots that can run, dance and perform backflips, manufacturers are now focusing on the software behind those machines. The goal is to help robots understand their surroundings, learn unfamiliar tasks and adapt without requiring individual programming for every job.
This emerging field is known as embodied AI. It combines artificial intelligence with physical machines, allowing robots to interact with the real world.
According to research firm ITjuzi, Chinese embodied AI companies raised 93.5 billion yuan, or about $13 billion, across 322 deals in the first half of 2026. That figure was five times higher than the amount raised during the same period a year earlier.
As competition pushes hardware prices lower, investors and manufacturers are increasingly focusing on robot intelligence.
“Robot makers have concluded that the margin won’t come from the hardware but from the brain,” said Dermot McGrath, founder of Shanghai-based technology consultancy ZenGen Labs.
The growing sector is also creating another front in the US-China technology rivalry. American companies lead in some advanced robotics AI systems. However, Chinese developers believe their manufacturing capacity and access to real-world training data could help them compete.
The country that successfully combines AI with robotics could gain an advantage in factory automation and logistics. In turn, this could reshape industrial productivity and the global robotics market.
“The brain is indeed the weakest link in the complete robotics stack,” said Spirit AI co-founder Gao Yang.
American startups, including Physical Intelligence, Skild AI and Generalist, are developing software that allows robots to learn different skills without extensive retraining.
Generalist says its GEN-1.5 model completed around 60% of simple tasks after watching demonstrations lasting between three and 12 seconds.
Meanwhile, Chinese company X Square has published research showing robots learning object-handling skills from a single human video. The approach does not require retraining the model for each task.
National CERT Issues Urgent Warning Over TeamViewer Security Vulnerability
Robotics consultant Georg Stieler said leading American laboratories were setting a demanding benchmark for how quickly robots could learn and reuse skills. However, he also considers China’s strongest startups credible competitors.
The two countries have different advantages. US developers benefit from advanced AI research and greater access to powerful computing chips.
By contrast, Chinese companies can work closely with robot manufacturers and industrial customers. This allows them to test and improve their systems in real working environments.
Specialist Chinese startups, including X Square, Galaxea and Spirit AI, have each reached valuations of roughly $3 billion, according to company disclosures.
“We want robots to shift from being driven by code to being driven by common sense,” said Yang Qian, co-founder and chief operating officer of X Square.
Established manufacturers are also increasing their investment in robot intelligence.
Hong Kong-listed UBTech advertised a chief scientist position overseeing AI model development. The advertised compensation package could reach 124 million yuan, or about $18 million, including equity and benefits.
Similarly, Hangzhou-based Unitree has around 30 researchers developing AI that predicts how objects behave, according to a person briefed by the company. Unitree is among the world’s largest humanoid robot makers by shipments.
Unitree did not respond to a request for comment.
Unlike chatbots, which learn from vast collections of online text, robots need examples of physical interactions.
These include gripping objects, handling flexible wires and completing other tasks that require movement and coordination.
Consequently, Chinese companies are exploring different methods of collecting training data. Some employ people to demonstrate tasks, while others rely on computer simulations. Several companies combine both approaches.
Spirit AI has hired around 1,000 people to record everyday and industrial activities, Gao said.
Meanwhile, ACE Robotics uses wearable sensors to capture workers’ movements and grip forces in shops and hotels.
Ant Group’s robotics unit, Robbyant, said it trained a model using 160,000 hours of material. This included 50,000 hours of human-guided robot movements.
The company also trained its software across 20 robot configurations from 17 brands. Its aim is to develop AI that can operate across different machines.
By comparison, DexForce uses computer-generated environments to train robots.
Chief operating officer Nicole Yang said the company can prepare some industrial tasks for deployment within six to eight hours. She also reported success rates above 99.9% in specified precision operations.
However, simulations have limitations. Jerry Wang of spatial intelligence company Manycore Tech said computer-generated environments cannot perfectly reproduce physical interactions.
X Square’s Yang also stressed the importance of training data quality.
“Data quality is more important than data diversity, and more important than simply collecting data at massive scale,” he said.
Despite growing investment, Chinese embodied AI companies face several challenges.
One major concern is limited access to advanced Nvidia chips because of US export controls. At the same time, regulatory concerns over inflated valuations and overcrowded markets are putting pressure on IPO funding.
Zhu Mingxuan, who leads development of JoyIn Technology’s Aether robot AI, warned that shortages of advanced chips could become a serious problem for the industry.
As robot AI systems become more powerful, developers will need greater computing capacity. Therefore, access to advanced hardware could influence how quickly companies improve their models.
Chinese developers are already testing whether their software can help robots solve unfamiliar problems.
Zhu described an experiment involving a robot that repeatedly failed to place headphones inside a cupboard beyond its reach.
Engineers then showed the robot videos of people completing the task. After watching them, the robot pushed a box towards the cupboard and tried to step onto it, Zhu said.
“Every time it was exploring a different behaviour,” she said.
Meanwhile, some companies are moving beyond laboratory demonstrations and testing robots in commercial settings.
Battery maker CATL said Spirit AI robots connect flexible wires during battery-pack testing at its Zhongzhou plant. The company reported an insertion success rate above 99%.
Similarly, Robbyant said its robots handle orders at two pharmacies in Shanghai. Staff members check the medicines before dispatch.
These deployments show how companies are testing robot AI in specific industrial and commercial tasks.
However, the larger goal is to develop software that can learn new jobs without extensive adaptation. Achieving that will require models that perform reliably across different customers, workplaces and machines.
Chinese developers are trying to reduce computing demands through more efficient models, synthetic data and simulations. Meanwhile, domestic chipmakers, including Huawei, are developing alternatives to Nvidia hardware.
Nevertheless, it remains unclear how far these efforts can offset the shortage of advanced chips.
Unitree founder Wang Xingxing said last month that the company had tested OpenAI’s GPT-6 on robots, with encouraging results.
He also suggested that GPT-7, if released next year, could bring robotics its own “ChatGPT moment”. The comparison refers to the major shift in public and commercial interest that followed ChatGPT’s launch in 2022.
Wang said he would consider that milestone reached when robots could complete 80% of tasks in 80% of unfamiliar environments using spoken instructions.
However, robotics consultant Georg Stieler cautioned that successful demonstrations would not automatically make robots reliable enough for widespread commercial use.
“If the question is when a model first handles a broad range of unfamiliar tasks in demonstrations, the optimists may well be right,” he said.
“If the question is when a customer can count on it, I would add years, not months.”
His comments highlight the industry’s central challenge: turning impressive demonstrations into dependable machines that can perform a wide range of real-world jobs.
[…] China Invests Billions in Embodied AI as Robotics Race Shifts Beyond Hardware […]