After years of building humanoid robots that can run, dance and perform acrobatics, Chinese companies are turning their attention to a different challenge: developing AI that can make robots useful across a wider range of jobs. The shift is driving investment in embodied AI—software designed to help machines interpret their surroundings, learn new tasks and adapt without being programmed for each one individually.
Chinese embodied-AI companies raised 93.5 billion yuan in 322 deals during the first half of 2026, five times the amount recorded a year earlier, according to data research firm ITjuzi. Funding is increasingly going to developers of robotic intelligence as well as hardware manufacturers. As competition pushes down robot prices, companies are looking to software for future profits. “The brain is, in fact, the weakest link in the whole robotics stack,” said Gao Yang, co-founder of Spirit AI.
The emerging sector is opening another front in the technology rivalry between China and the United States. US companies lead in some advanced robotics AI systems, while Chinese developers are counting on their manufacturing capacity and access to real-world training data to compete. The ability to combine AI and machines could influence automation in factories and logistics, although analysts say a breakthrough in demonstrations does not necessarily mean robots are ready for widespread commercial use.
US startups including Physical Intelligence, Skild AI and Generalist are developing software intended to teach robots different skills without extensive additional training. Generalist says its GEN-1.5 model completed about 60% of simple tasks after watching a demonstration lasting between three and 12 seconds. Chinese company X Square has published research showing robots learning manipulation skills from a single video of a person, without retraining the model for each task.
Chinese firms are also working closely with robot makers and industrial customers to test and improve their systems. X Square, Galaxea and Spirit AI have each disclosed valuations of about $3 billion. Established manufacturers are investing too: Hong Kong-listed UBTech advertised for a chief scientist to develop AI models, with compensation of up to 124 million yuan including equity and benefits. Unitree has about 30 researchers working on AI that can predict how objects will behave, according to a source familiar with the company.
Training these systems requires examples of physical interactions, such as grasping objects or handling flexible wires—not just the large collections of online text used to train chatbots. Chinese companies are gathering data in several ways, including human demonstrations and computer simulations. Spirit AI has hired about 1,000 people to record everyday and industrial activities, Gao said. ACE Robotics uses wearable sensors to capture workers’ movements and grip strength in shops and hotels.
Robbyant, the robotics unit of Ant Group, said it trained a model on 160,000 hours of material, including 50,000 hours of human-guided robot movements. The company also trained its software on 20 robot configurations from 17 brands, aiming to use the same AI across different machines. DexForce, meanwhile, uses computer-generated environments. Its chief operating officer, Nicole Yang, said some industrial tasks can be prepared for execution in six to eight hours, with success rates above 99.9% in specific precision operations. However, simulations cannot perfectly reproduce physical interactions, said Jerry Wang of simulation-environment designer Manycore Tech.
Companies are also testing whether their software can help robots respond to unfamiliar problems. Zhu Mingxuan, who leads development of JoyIn Technology’s Aether robotics AI, described a robot that repeatedly failed to place headphones in a cabinet beyond its reach. After engineers showed it videos of people doing the task, the robot pushed a box toward the cabinet and tried to climb onto it. “Each time it happened, it explored a different behavior,” Zhu said.
Some systems are already being used in real workplaces. CATL said Spirit AI robots connect flexible wires during battery testing at its Zhongzhou factory, with an insertion success rate above 99%. Robbyant said its robots process orders at two pharmacies in Shanghai, where staff check the medicines before dispatch.
The longer-term goal is software that can learn new jobs without major adjustments and perform reliably across different customers and work environments. Chinese companies are trying to reduce their computing needs with more efficient models, synthetic data and simulation, while domestic chipmakers including Huawei are developing alternatives to Nvidia hardware. But it remains unclear how far those efforts can offset a shortage of advanced chips, which Zhu said could become a serious problem for China’s embodied-AI industry.
Unitree founder Wang Xingxing said last month that the company had tested OpenAI’s GPT-6 on robots with encouraging results. He suggested that GPT-7, if released next year, could bring robotics its own “ChatGPT moment.” Wang said he would consider that threshold reached when robots could complete 80% of tasks in 80% of unfamiliar environments based on spoken instructions. Robotics consultant Georg Stieler cautioned that demonstrations and dependable commercial deployments are different milestones: a customer being able to rely on such systems could take years, rather than months.
Sources: CNN Brasil.