China’s robotics industry is shifting its biggest bet from machines that run, dance and backflip to software that can help them handle unfamiliar tasks. The ambition is to move humanoid robots beyond carefully staged demonstrations and into factories, warehouses and service jobs, where a robot must do more than repeat a programmed sequence.
The money is following that shift. Chinese embodied-AI companies raised 93.5 billion yuan, or about $13 billion, across 322 deals in the first half of 2026, five times the amount raised a year earlier, according to research firm ITjuzi. The emerging contest is not simply over who can build the most convincing humanoid. It is over who can develop the “brain” that makes the hardware commercially useful.
That race is also becoming a new front in US-China technology competition. US developers have advantages in advanced AI research and access to powerful computing chips. China’s wager is that its manufacturing base, industrial customers and access to real-world training environments can help close the gap.
The next breakthrough is not another backflip
The field, known as embodied AI, combines artificial intelligence with machines that perceive and interact with the physical world. Unlike chatbots trained on enormous volumes of text, robots need data showing how objects behave: how a hand grips a cup, how a cable bends or how a machine should respond when an object is out of reach.
“The brain is indeed the weakest link in the complete robotics stack,” said Gao Yang, co-founder of Spirit AI.
The challenge is to develop systems that can transfer a skill from one task or environment to another, rather than requiring engineers to program every movement separately. American startup Generalist says its GEN-1.5 model completed about 60% of simple tasks after watching demonstrations lasting between three and 12 seconds. China’s X Square has published research showing robots learning handling skills from a single human video without retraining the model for each task.
“The leading American labs are setting a demanding benchmark, particularly in how quickly robots can acquire and reuse skills,” said robotics consultant Georg Stieler, who considers China’s strongest startups credible competitors.
China is building a data pipeline for machines
Chinese developers are using different methods to teach robots physical skills. Spirit AI has hired about 1,000 people to record everyday and industrial activities. ACE Robotics uses wearable sensors to capture workers’ movements and grip forces in settings such as stores and hotels.
Ant Group’s robotics unit, Robbyant, says it trained a model using 160,000 hours of material, including 50,000 hours of human-guided robot movements. Its software was trained across 20 robot configurations from 17 brands, with the aim of running the same AI on different machines.
Other firms are leaning on simulation. DexForce says it can prepare some industrial tasks for deployment within six to eight hours and has reported success rates above 99.9% in specified precision operations. But simulated environments cannot perfectly reproduce real-world physical interactions, according to Jerry Wang of Manycore Tech, which develops simulated environments for AI training.
The issue is not simply how much data companies collect, but whether it teaches robots to cope with conditions they have not encountered before.
“Data quality is more important than data diversity, and more important than simply collecting data at massive scale,” said X Square co-founder and COO Yang Qian.
From a demonstration to a job that pays
The commercial test is already underway, though deployments remain narrow. Battery maker CATL said Spirit AI robots connect flexible wires during battery-pack testing at its Zhongzhou plant, with insertion success above 99%. Robbyant said its robots handle orders in two Shanghai pharmacies, where staff check medicines before dispatch.
These examples show why manufacturers are interested in robot intelligence: the potential lies in automating tasks that require more than fixed, repetitive movements. But a robot that works in one factory or pharmacy is not necessarily ready to perform the same task at another site, with different equipment, layouts or objects.
One experiment described by JoyIn Technology illustrates the difference between scripted movement and adaptive behavior. A robot repeatedly failed to put headphones into a cupboard beyond its reach. After engineers showed it videos of people solving the problem, it pushed a box toward the cupboard and tried to step onto it.
“Every time it was exploring a different behaviour,” said Zhu Mingxuan, who leads development of JoyIn’s Aether robot AI.
The goal is to make that kind of adaptation dependable, not merely an occasional success in a controlled test.
The chip constraint could limit China’s ambition
The investment surge has also raised the stakes for Chinese robotics firms. Specialist startups X Square, Galaxea and Spirit AI have each reached valuations of roughly $3 billion, according to company disclosures. Established manufacturers are competing for talent as well: UBTech advertised a chief scientist role overseeing AI model development with compensation of up to 124 million yuan, or $18 million, including equity and benefits.
But advanced robot intelligence requires substantial computing power. Chinese companies face shortages of advanced Nvidia chips because of US export controls, while concerns about crowded valuations and inflated expectations are squeezing IPO funding. Firms are trying to reduce computing needs through more efficient models, synthetic data and simulation; domestic chipmakers, including Huawei, are developing alternatives to Nvidia hardware. How far those efforts can offset the constraints remains uncertain.
What would count as robotics’ “ChatGPT moment”?
Unitree founder Wang Xingxing has suggested that GPT-7, if released next year, could help bring robotics its own “ChatGPT moment.” He defined the threshold as robots completing 80% of tasks in 80% of unfamiliar environments when given spoken instructions.
That is a demanding benchmark, and not the same as proving commercial reliability. A robot may perform well in a demonstration yet fail when a customer depends on it every day, across different settings and tasks.
“If the question is when a model first handles a broad range of unfamiliar tasks in demonstrations, the optimists may well be right,” Stieler said. “If the question is when a customer can count on it, I would add years, not months.”
For China’s robotics industry, the race has moved beyond making machines look capable. The next phase will be judged by whether they can learn, adapt and deliver consistent work, and whether companies can make that intelligence affordable at scale.
Shelton Rozario