China’s robot makers have mastered movement. Now comes the race for the brain

After years of building humanoid robots that can run, dance and do backflips, Chinese companies are now directing resources toward AI that could make these machines useful for a much wider range of work. The goal is software that lets robots understand their surroundings, learn unfamiliar tasks and adapt without being programmed for each job.

According to ITjuzi, Chinese AI companies raised 93.5 billion yuan ($13 billion) across 322 deals in the first half of 2026, five times the amount a year earlier. Funding is increasingly reaching specialized developers of robot intelligence as well as manufacturers. Dermot McGrath of ZenGen Labs says that as hardware prices fall, robot makers have concluded that profit will come “from the brain, not the hardware.” Spirit AI co-founder Gao Yang also calls the brain the weakest link in the whole robotics stack.

The race for intelligence

American startups such as Physical Intelligence, Skild AI and Generalist are building software that teaches robots varied skills without extensive training. Generalist says its GEN-1.5 model completed about 60% of simple tasks after watching a demonstration lasting 3 to 12 seconds. China’s X Square has published research showing robots acquiring control skills from a single human video without retraining the model.

US developers benefit from cutting-edge research and access to powerful chips, while Chinese firms can work closely with robot manufacturers and industrial customers to test and refine their systems. X Square, Galaxea and Spirit AI are each valued at roughly $3 billion. Established manufacturers are investing too: UBTech advertised a chief scientist role overseeing AI model development with compensation of up to 124 million yuan, and Unitree has about 30 researchers working on AI that predicts how objects behave.

The race for training data

Unlike chatbots, which learn from vast amounts of internet text, robots need examples of physical interaction. Companies are taking different routes. Spirit AI has hired about 1,000 people to record everyday and industrial activities, ACE Robotics uses wearable sensors, and Ant Group’s Robbyant unit built a model on 160,000 hours of material, including 50,000 hours of human-operated robot movements. DexForce favors simulation: it says it can prepare some industrial tasks for deployment in 6–8 hours, with success rates above 99.9% on certain precise operations. Jerry Wang of Manycore Tech points out that simulations cannot perfectly reproduce physical interaction, while Yang Qian of X Square says data quality matters more than scale or diversity.

Chinese firms face other obstacles as well: a shortage of advanced Nvidia chips due to US export controls, and tight IPO funding amid regulatory concerns about inflated valuations and overcrowding. As robots become more capable, the chip shortage will become a more serious problem. Companies are trying to cut compute needs through more efficient models, synthetic data and simulation, while domestic chipmakers such as Huawei develop alternatives to Nvidia. How far this will close the gap remains unclear.

From training to work

Zhu Mingxuan, who leads development of JoyIn’s Aether model, described an experiment in which a robot repeatedly failed to place headphones on an unreachable shelf. After engineers showed it videos of people doing the task, it pushed a box toward the shelf and tried to climb onto it. Real-world deployments leave no room for improvisation: CATL says Spirit AI robots connected flexible cables during battery pack tests at its Zhongzhou plant with a success rate above 99%, and Robbyant says its robots take orders at two pharmacies in Shanghai, with staff checking the medicines before they are sent out. The long-term prize is software that can learn new jobs without extensive adaptation, which will require training ever more powerful models on advanced chips.

The “ChatGPT moment”

Unitree founder Wang Xingxing said last month that the company has been testing OpenAI’s GPT-6 on robots, with promising results. If GPT-7 is released next year, he said, robotics could have its own “ChatGPT moment,” a capability shift like ChatGPT’s 2022 launch. He says the threshold will be reached when robots can complete 80% of tasks in unfamiliar environments from spoken instructions.

Consultant Georg Stieler cautions that progress in demonstrations does not mean robots are ready for broad commercial use. If the question is when a model will first perform a wide range of unfamiliar tasks in demos, the optimists may be right, he says. If the question is when a customer can rely on it, he would add years, not months.