Play store
Humanoid robots could achieve a breakthrough in their intelligence by late 2027, comparable to the impact ChatGPT had on artificial intelligence, according to Wang Xiaogang, CEO of Chinese embodied AI startup ACE Robotics.
“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 told Reuters on Friday.
However, he said it could take another four to five years after that breakthrough for embodied AI models to see widespread commercial use across industries.
While large language models such as ChatGPT and DeepSeek have become widely used, robots capable of performing a broad range of tasks autonomously in unfamiliar physical environments remain some way off.
Embodied AI models are designed to give robots the ability to understand, navigate and interact with the physical world in real time.
Unlike conventional language models, they combine perception, multimodal understanding, physical simulation and action planning.
Wang Xingxing, founder of Chinese robotics company Unitree, also predicted this week that robot intelligence could see a major breakthrough within two to three years at the earliest. However, robotics executives say access to high-quality real-world training data remains a major challenge.
ACE Robotics, founded in July 2025 and backed by Ant Group and SenseTime, has raised more than $100 million through several funding rounds in the first half of 2026. Wang said the company aims to pursue an initial public offering “as early as permitted.”
The company’s open-source Kairos-4B model is currently ranked highly in public benchmarks, despite having only 4 billion parameters. Its world model can generate long-horizon predictions of video and robot actions over several minutes.
ACE is also expanding real-world data collection by equipping workers on production lines with lightweight sensors. Wang said the company expects to accumulate tens of millions of hours of training data within two years.
Many humanoid-robot developers currently rely on physical teleoperation, in which workers wearing exoskeletons and controllers repeatedly perform physical movements to generate training data.