China's Robot Makers Have Mastered the Moves. Now Comes the Race for Brains

Chinese companies have spent years building robots that can run, dance, and even do backflips. Now those same companies are focused on a harder problem: giving robots a brain. They are investing billions of dollars into a kind of software called 'embodied AI' that lets robots figure out new tasks on their own, without being programmed for every single job.

In the first half of 2026, Chinese companies raised about $13 billion for embodied AI across more than 300 deals. That is five times more money than they raised in the same period a year earlier. The money is going not just to robot builders, but also to specialist companies focused entirely on making robots smarter.

Hardware, meaning the physical parts of a robot, is getting cheaper as more companies compete to build it. Because of this, robot makers now believe the real value will come from the software that runs the machine. As one expert put it, the margin won't come from the hardware but from 'the brain.' This race for robot intelligence is now a major part of the rivalry between the United States and China in technology.

American startups like Physical Intelligence, Skild AI, and Generalist are building software to teach robots new skills quickly. Generalist says its latest AI model can complete about 60% of simple tasks after watching a short demonstration of just 3 to 12 seconds. A Chinese company called X Square has shown that its robots can learn to handle objects just by watching a single human video, without needing to be retrained each time.

Chinese startups like X Square, Galaxea, and Spirit AI have each reached values of around $3 billion. Big manufacturers are also joining the race. Hong Kong-listed UBTech posted a job for a chief scientist with pay of up to $18 million, including stock. Unitree, one of the world's biggest humanoid robot makers, has around 30 researchers working on AI that helps robots predict how objects will move and behave.

One of the biggest challenges is training data — the examples robots need to learn from. Unlike chatbots that learn from text on the internet, robots need examples of real physical actions, like picking up objects or handling wires. Chinese companies are tackling this in different ways, from hiring people to demonstrate tasks, to using computer simulations, to combining both methods.

Spirit AI hired about 1,000 people to record everyday and industrial tasks. Ant Group's robotics unit, Robbyant, trained its AI using 160,000 hours of material, including 50,000 hours of robot movements guided by humans. Another company, DexForce, uses computer-generated environments to train its robots and says it can get some industrial tasks ready in just 6 to 8 hours, with success rates above 99.9%.

Chinese robot companies also face serious obstacles. The United States has placed export controls on advanced computer chips made by Nvidia, making them hard for Chinese companies to get. Without powerful chips, it becomes much harder to train more capable AI, and one developer warned this chip shortage will become a serious problem for the whole industry.

Despite these challenges, some robots are already being used in the real world. Battery maker CATL said robots from Spirit AI are connecting flexible wires during battery testing, with a success rate above 99%. Robbyant said its robots help handle medicine orders in two pharmacies in Shanghai, showing that robot AI is moving beyond the lab and into actual workplaces.

The longer-term goal is software that can learn brand-new jobs with very little extra training. Unitree's founder said that if a powerful enough AI model is released next year, it could be a major turning point for robotics — similar to how ChatGPT's launch in 2022 made AI a mainstream tool for millions of people. He said he would consider that milestone reached when robots could complete 80% of tasks in 80% of unfamiliar environments using spoken instructions alone.

Not everyone is so optimistic about the timeline. One industry consultant said a big leap in robot performance during demonstrations does not mean robots are ready for wide commercial use. 'If the question is when a customer can count on it, I would add years, not months,' he said. Still, the rapid progress in both China and the US suggests that smarter, more adaptable robots may not be too far away.