Spirit AI plans to create a breakthrough brain for humanoid robots by 2027
English translation from Unibot. Source published 2026-09-21.
China's Spirit AI promises to train the brains of its humanoid robots to understand commands in natural language and take on a wide range of physical tasks by mid-2027. The company's co-founder and chief scientist, Gao Yang, told Reuters about this.
True, Gao himself set the bar more modestly than the agency. Reuters writes about “most general-purpose tasks,” but the scientist only says that the robot will hear the request and perform a chain of intelligent actions in response.
Prediction, not measured result
Today, Spirit AI claims a 90% success rate on simple tasks in furnished living rooms. But Reuters does not provide either the test methodology, the number of attempts, or the definition of “success.”
Gao himself admits that fine motor skills are still poor - unscrewing the cap of a bottle does not always work, and the models cope poorly with unfamiliar tasks.
Factories and money
Several dozen Moz1 wheeled humanoids are already working on the production lines of battery manufacturer CATL and retailer JD.com. JD.com, by the way, is also an investor.
But what exactly the robots do on the line and how autonomously they do is not stated in the report. So you can’t draw conclusions based on deployment figures alone.
The company has 300 people. Since its founding in 2024, it has raised more than $670 million. The current valuation is 20 billion yuan, approximately $3 billion. Gao refused to talk about plans to enter an IPO.
When to wait in everyday life
The next year or two, he said, will open the first window for industrial use. Simple tasks in the service sector will be improved within two years. It will come home even later.
Gao estimates that it takes at least eight years for a humanoid to become useful in everyday life. Unstructured space and unpredictable interactions raise the technical bar too high.
Real world data
About a thousand contractors across China collect data on human movements for Spirit AI - in apartments and factories. At a Beijing training center visited by Reuters, workers wearing sensors endlessly opened refrigerators, opened safes and cut vegetables with knives.
The company focuses on real-world interactions rather than simulations. Gao explains: simulators are good at describing rigid objects, but fall short of flexible ones - say, an electrical cable.
Spirit AI also prefers diverse and imperfect records, which it calls “dirty data.” According to Gao, this variety speeds up model training better than narrowly defined pure movements - which at some other Chinese centers require more than 50 repetitions.
Safety and force control
The whole body force control system is responsible for physical safety: if the interaction force exceeds a preset limit, Moz1 automatically brakes.
Gao calls this the basic solution. More advanced alignment research, he says, will become practical when robot foundation models gain much more autonomy.