"We expect to reach the 'ChatGPT moment' for embodied intelligence by the end of next year, driven by world models and environmental data capture," Wang Xiaogang told Reuters.

"Even if we reach that inflection point by late 2027, it will likely take another four to five years to see broad commercial implementation of embodied world ​models across sectors," said Wang, who is also a co-founder of Chinese AI visual recognition pioneer SenseTime.

While large language models such as ChatGPT ​and DeepSeek have become a staple in workplaces and households globally, AI models that allow a robot to smoothly ⁠complete a wide range of tasks in unfamiliar physical environments remain distant.

And as more companies in China's fledgling humanoid robot industry seek funding and high ​valuations, investors are placing more importance on real-world deployment over activities such as dance or athletics to demonstrate their economic value.

Embodied AI models determine the intelligence and ​autonomous operation abilities of robots. Unlike large language models, these physical AI simulation systems are designed to help robots understand and navigate real-world environments in real time.

Wang Xingxing, the founder of China's Unitree, this week predicted that robot brains could see a dramatic breakthrough in two to three years at the earliest. He and other Chinese robotics CEOs have acknowledged ​that acquiring high-quality real-life training data for these models remains a bottleneck.

ACE Robotics, which was founded in July 2025 and is backed by Ant Group ​and SenseTime, has raised more than $100 million in the first half of this year through several financing rounds.

Wang said it aims to launch an initial public offering (IPO) "as early as ‌permitted". Chinese ⁠listing rules typically require companies to have at least three fiscal years of operation.

The company's open-source Kairos-4B model is ranked first globally by public benchmarks, outperforming world models such as Nvidia's Cosmos 3 and Ant Group's Lingbot despite having a much smaller 4 billion parameter count.

The world model integrates perception, multi-modal understanding, physical simulation and action planning. It can also generate long-horizon video and action predictions over multi-minute sequences.

EXPANDING TRAINING DATA

Many robot companies including Unitree and startup X Square ​are developing their own embodied AI ​foundation models.

"Over the past few years, ⁠the entire industry has accumulated data of roughly 100,000 hours, which is far from enough to train embodied foundation models," said ACE's Wang.

He said ACE is rapidly scaling data collection by using people on real production lines fitted with ​lightweight sensors.

"The collection efficiency is extremely high ... We expect to accumulate tens of millions of hours of data ​within two years," Wang ⁠added.

Many firms train humanoid robots using physical tele-operation, where workers wearing exoskeletons and controllers repeat simple physical movements hundreds of times a day.

ACE is deploying its embodied AI models in commercial settings such as unmanned retail stores, hotel services and instant-delivery warehouses staffed by humanoid robots from Chinese makers including Unitree, AgiBot and Fourier.

"We plan ⁠to deploy ​in at least 1,000 stores over the coming year, scaling to 10,000 stores in two ​years," said Wang.

For training its models, ACE uses AI chips from Nvidia as well as Chinese makers including Rhino Tech and Digua Robotics.

"This diversifies our supply chain, because in the future, we will ​definitely need comprehensive intelligent hardware solutions, and their costs need to be significantly reduced," said Wang.

(This story has been corrected to fix Wang Xiaogang's title to chairman, not CEO, of ACE Robotics, in the headline and paragraph 1)