When Will Humanoid Robots Have Their ChatGPT Moment?

The two timelines

Wang Xiaogang: a technical breakthrough by late 2027

Wang Xiaogang said ACE expects embodied intelligence to reach its “ChatGPT moment” by the end of 2027, driven by better world models and much larger collections of environmental data. He separated that technical breakthrough from broad commercial deployment, which he said could require another four to five years.

That distinction matters. A model may demonstrate a major improvement in reasoning or control while still being too expensive, inconsistent, or difficult to integrate into everyday operations. A research milestone and an economically dependable workforce are not the same event.

Wang Xingxing: two to three years in the best case, up to 10 years otherwise

Wang Xingxing described two possible paths. In the optimistic scenario, embodied intelligence could reach its tipping point in two to three years. If progress in generalization and precise physical control slows, he said the industry could instead need five to 10 years.

His household-task benchmark also makes the forecast more demanding than a simple capability demo. The robot must operate in a new environment, interpret natural-language instructions, and complete a broad range of tasks autonomously.

Why robot “brains” are the bottleneck

The hardware is improving quickly, but the software still struggles to generalize from limited experience to the messy variety of real-world situations. Both executives identified AI models and training data as the central constraint.

Robots need data that connects what they see with what they do: how an object feels, how much force to apply, what changes when a surface is slippery, and how to recover when an action fails. Simulation and demonstrations can help, but they do not fully capture the unpredictability of homes, factories, stores, and other physical environments.

This creates a data problem that differs from training a text model. A robot cannot simply learn from descriptions of an action; it must learn the relationship between perception, movement, consequences, and correction.

How ACE Robotics is trying to build that data pipeline

ACE Robotics is pursuing a “human-centered” approach: equipping production-line workers with lightweight sensors to capture human actions and environmental interactions. The company says it aims to collect tens of millions of hours of training data for its embodied world models.

The strategy is intended to expand the supply of real-world examples beyond data generated only by robots or simulations. Human operators can demonstrate how tasks are completed, how movements adapt to changing conditions, and how actions are chained together over longer workflows.

ACE was established in July 2025 and is led by Wang Xiaogang, a co-founder of SenseTime. Ant Group led its angel financing round, while SenseTime-related capital and other investors also participated.

What is notable about ACE’s Kairos model?

ACE says its open-source Kairos-4B embodied model led public embodied-AI benchmarks and outperformed Nvidia’s Cosmos 3 in the comparison cited at the conference. Those are company-linked performance claims, so they should be read as benchmark results rather than proof that general-purpose humanoid robots are ready for homes or factories.

The company is also pursuing deployment beyond laboratory demonstrations. ACE has said it plans to place its AI models in 1,000 unmanned retail stores. That ambition shows how developers are targeting constrained commercial environments first, where workflows can be standardized more easily than in a typical household.

China’s robotics boom adds pressure—and caution

The forecasts arrived amid intense enthusiasm for China’s humanoid-robot sector. Unitree’s Shanghai STAR Market IPO raised about $904 million and was reported to be more than 8,000 times oversubscribed by retail investors. Its shares closed 460% above the IPO price, giving the company a market value of roughly $50 billion.

Shipment data also shows China’s early manufacturing lead. Smart Analytics Global estimated that global humanoid-robot shipments reached about 19,100 units in the first half of 2026, with Chinese manufacturers accounting for more than 97% of the total.

That figure should not be confused with proof of mature, general-purpose automation. The available reporting contains a conflicting claim that China shipped more than 40,000 humanoid robots in the same period, but that number is inconsistent with the 19,100-unit global estimate and is not supported by the stronger shipment data available here.

Morgan Stanley has separately raised its forecast for China’s full-year humanoid-robot shipments to 50,000 units, citing faster-than-expected movement from demonstrations toward commercial deployment. A shipment forecast, however, measures market expansion—not whether the machines can pass the 80%-of-household-tasks test described by Wang Xingxing.

The real takeaway: capability may arrive before reliability

The most useful reading of the two executives’ comments is not that humanoid robots are guaranteed to reach a specific milestone in 2027. It is that the industry is approaching a point where software quality, training data, and generalization—not just motors, batteries, or mechanical design—will determine whether the category becomes broadly useful.

ACE’s forecast points to a technical breakthrough by late 2027 followed by years of commercialization. Unitree’s range makes the uncertainty explicit: two to three years if progress accelerates, five to 10 years if the hardest problems remain stubborn.

China is already showing strength in financing, manufacturing, and early shipments. The unresolved question is whether that scale can produce robots that work reliably in environments they have never seen. Until the data and model problem is solved, a “ChatGPT moment” for robot brains remains a forecast—not a finished product.