Wang Xiaogang, chairman of China’s ACE Robotics, anticipates a breakthrough in the field of embodied AI by the end of 2027. The key drivers will be more advanced robot control models and a significant expansion of datasets on real-world physical interactions, he told Reuters.
“We expect a ‘ChatGPT moment’ for embodied intelligence by the end of next year,” Wang said.
However, he estimates that widespread commercial adoption of the technology will take another four to five years.
Robots Lack ‘Brains’
Modern humanoids can already run, dance, and perform individual programmed tasks. However, they struggle with a wide range of tasks in unfamiliar environments.
At ACE Robotics, the main limitation is considered to be AI models rather than hardware. Embodied intelligence systems need to perceive the world in real-time, understand instructions, predict the consequences of their actions, and choose subsequent movements.
To address this, the startup is developing the Kairos model. It integrates perception, multimodal understanding, physical environment modeling, and behavior planning.
In July, the company announced that Kairos ranked first in four embodied AI tests, including tasks involving two-handed control, long-term planning, and physical interaction. According to Reuters, the open version Kairos-4B also outperformed Nvidia Cosmos 3 and Ant Group’s Lingbot in public benchmarks, despite having a comparatively small size of 4 billion parameters.
Data is the Main Shortfall
Another issue is the amount of information on real-world interactions, Wang noted. He said that the entire industry has collected about 100,000 hours of such data over the past few years. This is “far from enough” for training foundational embodied AI models.
Typically, companies gather information through teleoperation. A person in an exoskeleton or with controllers repeatedly performs the required movement, and the robot records the trajectory.
ACE Robotics uses a different approach. The company equips people already working on real production lines with lightweight sensors and records their actions. Over two years, the startup aims to accumulate tens of millions of hours of data.
ACE Targets 10,000 Stores
The firm is testing its models in areas such as unmanned retail, hotel services, and fast delivery warehouses. The hardware base consists of humanoid robots from Chinese manufacturers Unitree, AgiBot, and Fourier.
Within the next year, ACE plans to implement the technology in at least 1,000 stores, aiming to reach 10,000 by the end of the second year.
To train its models, the startup uses Nvidia accelerators, as well as Chinese chips from Rhino Tech and Digua Robotics. Wang explained this diversification as an effort to reduce equipment costs and dependency on individual suppliers.
ACE Robotics was founded in July 2025. Its investors include Ant Group and SenseTime. In the first half of 2026, the company raised over $100 million and plans to go public as soon as Chinese regulations allow.
In June, Morgan Stanley analysts raised their forecast for Chinese humanoid deliveries in 2026 from 28,000 to 50,000 units. They cited commercialization, government support, and large-scale deployments for data collection as reasons.
