For the past two years, the most visible progress in Chinese robotics has come from hardware. Humanoid machines built by domestic manufacturers can sprint, dance and perform backflips on command, and clips of those demonstrations travel around the world within hours. The software that decides whether a robot can actually do something useful has received far less attention, even though engineers describe it as the harder problem. That layer is known in the industry as embodied artificial intelligence.
Spirit AI, a Beijing company that develops what it calls robot brains, used a briefing at its offices to argue that this layer is both the weakest part of the robotics stack and the part closest to a sudden jump in capability. Gao Yang, the company's co-founder and chief scientist, set out a timeline that places a qualitative shift within reach in roughly a year, while pushing the household robot much further out.
Reuters reported on September 18, 2026 that Gao told the news agency humanoid robots will be able to complete most general-purpose tasks from verbal instructions as soon as next year, while machines that are genuinely useful around the house remain eight or more years away. "The brain is indeed the weakest link in the complete robotics stack," he said at Spirit AI's Beijing offices. The industry, he added, is searching for its ChatGPT moment, a reference to the model release that turned large language models from a research curiosity into a mass market tool.
Gao said the target is a GPT-3.0 style milestone by mid-2027, after which a person will be able to speak to a robot in natural language and watch it execute a series of reasonable physical actions in an attempt to complete the task. That claim is aggressive by the standards of most robotics researchers, but it is deliberately narrower than the company's ambitions for the home, where the timeline is measured in years rather than months.
Key Facts
The milestone Gao describes is specific. He told reporters that Spirit AI expects to reach what he calls GPT-3.0 level capability by mid-2027, meaning a robot can take a natural language instruction and carry out a sequence of logical physical actions in an attempt to finish the job. The comparison is to the generation of large language models that demonstrated, around 2020, that scaling data and computing power could produce systems with broad but unreliable general ability. Gao framed the robot brain as the weakest link in the entire robotics system, a phrase he repeated in more than one interview with visiting journalists.
Spirit AI's commercial footprint is still small. Reuters reported on September 18, 2026 that the company has tens of its own Moz1 wheeled humanoid robots deployed on production lines at the battery maker CATL and at the retailer JD.com, which is also an investor. The company employs around 300 people and has raised more than $670 million since it was founded in 2024. It is now valued at 20 billion yuan, or roughly $2.9 billion, placing it among the most rapidly capitalized firms in China's embodied intelligence sector. Gao declined to comment on any plans for an initial public offering, BNN Bloomberg reported on September 18, 2026.
Training data is collected by people rather than generated by computers. Spirit AI employs roughly 1,000 contractors across China who wear data collection equipment at home or in factories. During a visit to a training centre at the company's Beijing office, dozens of young people fitted with sensors repeated motions such as opening fridges, unlocking safes and cutting vegetables with knives. BNN Bloomberg reported on September 18, 2026 that the startup relies overwhelmingly on real world data instead of virtual simulations, with Gao noting that simulators handle rigid bodies well but that flexible objects such as deformable electric cables remain a problem.
The published performance numbers are modest but concrete. Mezha reported on September 18, 2026 that in structured living room environments Spirit AI's robots already complete simple tasks with a success rate of around 90 per cent. Tens of Moz1 wheeled units are working on CATL production lines and inside the JD.com retail network, which has also invested in the company. Difficulties remain in fine motor actions such as unscrewing a bottle cap and in handling tasks the models have never seen before. At other robot training facilities in China, operators may need to repeat a movement more than 50 times to get one clean sample, while Spirit AI has found that using what it calls dirty data, a broader and more varied range of motions, helps its models improve faster.
Analysis
The distance between the mid-2027 target and the eight year horizon for homes is the most revealing part of Gao's message. TV Delmarva reported on September 18, 2026 that Spirit AI points to a persistent data shortage as a key obstacle slowing the development of robot intelligence software. That framing places the bottleneck in the data pipeline rather than in chips, actuators or mechanical design, the areas where Chinese manufacturers have moved fastest and where the country's supply chains are strongest.
Gao's staged timeline is also a commercial strategy. He said the next one to two years mark the initial window for industrial applications, and that two years from now robots will be deployed in commercial service settings doing simpler tasks, with entry into homes described as far harder than both. What this really means is that Spirit AI is sequencing its bets: factories first, then service work such as retail and logistics, and only much later the unstructured environment of a family home. The 90 per cent success rate the company reports is measured in structured living room settings, which is precisely what a real household is not.
Valuation provides a second lens. A company with around 300 employees, founded in 2024, with tens of robots on customer production lines, has been valued at 20 billion yuan, or about $2.9 billion. Investors are therefore pricing a future platform rather than present revenue. The bigger picture here is that capital in China's embodied intelligence sector is being allocated on the expectation that software, not hardware, will determine which companies survive the next phase of the market.
The data collection model carries its own risks. Roughly 1,000 contractors wearing sensors in homes and factories is an expensive, hard to scale approach compared with the synthetic data pipelines used elsewhere in machine learning. BNN Bloomberg reported on September 18, 2026 that Spirit AI overwhelmingly relies on real world data, and Gao's own caveat about deformable cables suggests the company has chosen a path that is slower to build but potentially more robust once deployed. If the GPT-3.0 style milestone arrives by mid-2027, that bet will look prescient. If it slips, the cost base remains.
Why It Matters
For the wider robotics sector, Spirit AI's timeline sets a public benchmark that rivals will now be measured against. Chinese humanoid makers have already proven they can build machines that move impressively. The open question has always been whether those machines can understand a task, break it into steps and recover when something goes wrong. Gao's claim that verbal instruction plus a chain of physical actions is achievable within about a year is a testable promise, and the industry's progress toward it will be visible to customers, investors and regulators alike.
The home timeline matters just as much. Eight or more years is a long way out, and it implicitly challenges the marketing of consumer robots that are already on sale in some markets. Mezha reported on September 18, 2026 that widespread household use may begin much later than the industrial rollout. For anyone waiting for a general purpose domestic helper, the message from one of China's leading robot brain developers is that the hardware will arrive long before the intelligence does.
Next Up
Spirit AI's near term focus is expansion inside factories and retail. With tens of Moz1 robots already working at CATL and JD.com, the company is likely to add deployments in commercial service settings over the next two years, the window Gao identified for simpler tasks. The company declined to discuss a possible initial public offering, so its next public milestone will probably be a technical one: evidence that its models can handle unseen tasks and fine motor work such as unscrewing a bottle cap.
Watching whether the GPT-3.0 style milestone lands by mid-2027 will therefore be the single most useful signal for the sector. A 300 person startup with more than $670 million raised and a $2.9 billion valuation has effectively set its own deadline. Everything the industry says about robot intelligence between now and then will be read against it.
Comments (0)
Log in or sign up to leave a comment.
No comments yet. Be the first to share your thoughts.