Beijing-based Spirit AI has set a target date for a proposed embodied-AI capability: converting a spoken request into physical actions a robot can reasonably attempt. In a September 18 report, Reuters said Gao Yang, the company’s co-founder and chief scientist, expects a “GPT-3.0 milestone” by mid-2027. This is an executive forecast, not a demonstrated capability or independently validated benchmark.
Gao Yang’s mid-2027 robot-brain forecast is a company expectation
Gao’s formulation is specific about the intended interaction. By mid-2027, he said, a person should be able to speak to a robot in natural language and have it execute “a series of reasonable physical actions to attempt the task.” It does not promise reliable completion of every request, general household operation, or human-equivalent understanding. Instead, it describes a proposed threshold for software translating language into actions.
Reuters defines embodied AI as the software layer that determines robot intelligence and enables economically productive physical tasks. It contrasts Chinese humanoid hardware advances—including sprinting, dancing, and backflips—with the software bottleneck. Reuters reported that Gao characterized the brain as the weakest link in the complete robotics stack. The forecast concerns the software layer rather than mechanical form alone.
(Related: Spirit AI Is Removed from RoboArena After Questions Over Benchmark Manipulation)
A 90% structured-room result does not establish household readiness
Spirit AI’s reported current result is much narrower than the mid-2027 ambition. Reuters said its robots achieved a 90% success rate for simple tasks in structured living-room environments. Gao called household deployment much harder and said it could take at least eight years. The figure does not establish performance in unfamiliar homes, complex domestic work, or general-purpose use.
Reuters reported that fine-motor work, including unscrewing a bottle cap, remains difficult, as do tasks the robot has not encountered before. Gao’s near-term waypoint therefore differs from his at-least-eight-year household estimate, which does not assign a fixed calendar year.
Spirit AI already has a narrower industrial footprint. Reuters reported that tens of Moz1 wheeled humanoid robots are on production lines at CATL and JD.com; JD.com is also an investor. Gao called the next one to two years “the initial window for industrial applications,” then projected simpler commercial-service tasks two years from now. Reuters presents these as Gao’s projections.
Moz1 deployments and safety controls define the nearer industrial case
Gao told Reuters that Spirit AI’s robots use whole-body force control and automatically trigger emergency braking when interaction force with the environment is excessive. He described that as a baseline safety policy. This is a company account of its safety approach, not an external safety assessment, but it identifies force management as part of the operating challenge.
The Spirit AI official product page describes Moz as having an “Intelligent Embodied Brain,” VLA technology, a 26-degree-of-freedom body, 7-DoF hybrid-controlled bionic arms, an omnidirectional wheeled base, and safety and collision detection. These vendor descriptions show what the company says it is building; they do not validate Reuters’ task result, forecast, or deployment report.
One thousand contractors supply real-world training data
Reuters reported that Spirit AI relies mainly on real-world data, rather than virtual simulations, to train robot brains. Roughly 1,000 contractors use wearable collection equipment in homes and factories. At its data center, sensor-equipped people repeat actions including opening refrigerators, unlocking safes, and cutting vegetables.
Gao said Spirit AI favors diverse “dirty data” over narrowly repeated “clean” movements and said it improved models faster. That is his account, not an established conclusion about robotic data. He added: “Simulators handle rigid bodies well, but flexible objects like deformable electric cables remain a problem.” Reuters notes competitors use simulators; the report does not establish a settled sector-wide data strategy.
Funding and architecture add context without validating the forecast
Reuters described Spirit AI as a 300-person startup that has raised more than $670 million since its 2024 founding and is valued at 20 billion yuan, or $2.9 billion. These reported funding, valuation, and staffing figures do not independently validate Gao’s proposed milestone.
In August 2024, Securities Times reported that Qianxun Intelligent, Spirit AI’s Chinese name, announced nearly 200 million yuan in seed and angel financing. It named Honghui Capital as angel-round lead; Dacheng Venture Capital and Qiancheng Capital as co-investors; and Shunwei Capital and Oasis Capital as continuing seed investors. This figure should not be combined with Reuters’ later total because the record does not reconcile them.
Securities Times reported a route combining a visual-language model for task decomposition, an embodied model for action generation, and model-predictive control for low-level motion. It identified Han Fengtao as founder and CEO and Gao as co-founder. That historical account does not independently corroborate the 2027 target.
In an August 20 Reuters context report, Unitree CEO Wang Xingxing said a major software leap could take two to three years optimistically or five to 10 years at the latest. That is not evidence for Spirit AI’s forecast.