The embodied artificial intelligence industry is on the brink of a tectonic shift. Wang Xiaogang, chairman of the Chinese startup ACE Robotics, is convinced that by the end of 2027, we will witness a kind of "ChatGPT moment" for robotics. According to him, the key catalyst will be a qualitative leap in control models and exponential growth in data on physical interaction with the world.
"We expect a 'ChatGPT moment' for embodied intelligence by the end of next year," Wang stated.
However, one should not expect an instant revolution in the market. The company's head predicts that widespread commercial adoption of such technologies will take at least another four to five years. This is a realistic horizon for overcoming infrastructure and regulatory barriers.
The problem is not the "hardware," but the "brains"
Modern humanoids already demonstrate impressive agility: they run, dance, and perform highly specialized tasks. But their main weakness is the inability to adapt to unfamiliar conditions and solve a wide range of non-standard problems. ACE Robotics rightly notes that the hardware side has long ceased to be the bottleneck. The main brake is the immaturity of AI models, which must process sensory information, interpret commands, and calculate the consequences of every movement in real time.
To solve this problem, the company is developing its own model, Kairos, which combines perception, multimodal understanding, and behavior planning. There are already successes: in July, Kairos topped four authoritative tests for embodied AI, including control of two manipulators and long-term planning. Notably, the open version, Kairos-4B, with only 4 billion parameters, outperformed heavier systems in public benchmarks, including Nvidia Cosmos 3 and Lingbot from Ant Group.
Data deficit — the main challenge
Wang Xiaogang emphasizes that the industry's main deficit is data on real physical interaction. Over all the years of development, the sector has collected only about 100,000 hours of such recordings, which is catastrophically insufficient for training foundation models. The traditional approach — teleoperation in exoskeletons — is too slow and expensive.
ACE Robotics proposes an innovative path: they equip ordinary workers on active production lines with lightweight sensors, recording their natural actions. This allows data to be collected on an industrial scale. The startup plans to accumulate tens of millions of hours within two years, which would become an undeniable competitive advantage.
Ambitious plans and scaling strategy
The company is already testing its solutions in unmanned retail, hotel service, and rapid-delivery warehouses. The hardware base is provided by humanoids from Unitree, AgiBot, and Fourier. The plans are impressive: within the next year, the technology will be deployed in at least 1,000 stores, and by the end of the second year — in 10,000. For training models, not only Nvidia is used, but also Chinese chips from Rhino Tech and Digua Robotics — a deliberate step to reduce costs and lessen dependence on a single supplier.
ACE Robotics, founded in July 2025 with support from Ant Group and SenseTime, has already raised over $100 million in the first half of 2026 and is preparing for an IPO as soon as local legislation permits.
My view: Wang's forecast looks ambitious, but not without foundation. The key factor is precisely data, and ACE Robotics' approach to collecting it in real-world conditions looks groundbreaking. If they manage to achieve the stated volumes, it could indeed accelerate the evolution of robots, turning them from highly specialized machines into universal assistants. However, the four-year lag to commercialization is an optimistic scenario, given the complexity of integration into existing business processes.