Chinese robotics firm says humanoid “brains” could see GPT style leap by 2027

Humanoid robot software could reach a breakthrough moment as early as mid 2027, but getting the machines into ordinary homes will likely take another eight years, according to the co-founder of Chinese robotics firm Spirit AI.

Gao Yang, who co-founded the Beijing based company and serves as its chief scientist, said the biggest obstacle facing the industry is no longer the physical hardware but the intelligence that controls it. “The brain is indeed the weakest link in the complete robotics stack,” he said in an interview at the firm’s Beijing offices. Gao, who also teaches robotics as an assistant professor at Tsinghua University, compared the coming shift to the arrival of OpenAI’s GPT 3.0 model, the system that eventually powered ChatGPT and reshaped the broader artificial intelligence industry.

Chinese humanoid manufacturers have already demonstrated striking physical capabilities, with robots able to sprint, dance and perform backflips on command. Gao said the industry’s attention has now shifted toward the software layer that determines how well a robot can think, adapt and carry out useful work in unpredictable, real world settings, since that layer ultimately decides how economically productive the machines become.

Spirit AI’s own robots have reached a 90 percent success rate on simple tasks inside structured living room style environments, Gao said. He laid out a staged timeline for the technology, describing the next one to two years as the initial window for industrial deployment, followed by a period in which robots move into commercial service roles handling simpler jobs. Entering private homes, he said, represents a far harder challenge than either of those steps and remains considerably further off.

The company currently operates tens of its wheeled Moz1 humanoid robots on production lines at battery manufacturer CATL and at e-commerce retailer JD.com, which is also among its investors. Founded in 2024, the 300 person startup has raised more than 670 million dollars, making it one of the most heavily funded firms in China’s embodied intelligence sector, with reports valuing the company at roughly 20 billion yuan, or about 2.9 billion dollars. Gao declined to comment on any plans for a public listing.

Gao pointed to data scarcity as the central bottleneck holding back progress on the software side. Rather than relying primarily on simulation to train its models, a method many rival firms use to cut training costs, Spirit AI leans heavily on real world data collection. Gao said simulated environments handle rigid objects reasonably well but still struggle with flexible items such as deformable cables, which remain a persistent problem for training accuracy.

To gather that real world data, the company employs roughly 1,000 contractors across China who wear motion capture equipment in homes and on factory floors. At the firm’s Beijing offices, a training centre houses workers fitted with sensors who repeat everyday actions such as opening refrigerators, unlocking safes and cutting vegetables with knives, generating movement data used to teach the robots those same tasks. Gao described this varied, imperfect data drawn from ordinary environments as more valuable for training than the cleaner data typically produced in simulation. In some Chinese training facilities, operators reportedly need to repeat a single movement more than fifty times to capture one clean recording that meets the precision required for the models.

Gao’s timeline sits alongside a range of predictions from other Chinese robotics leaders. The founder of rival firm ACE Robotics has forecast an equivalent breakthrough, which he termed the “ChatGPT moment” for embodied intelligence, arriving by the end of next year, driven by advances in world models and environmental data capture. Wang Xingxing, founder of the well known humanoid robot maker Unitree, has offered a more cautious estimate, placing the same milestone two to three years away at the earliest.

Gao said progress inside Spirit AI has already accelerated sharply since the company’s founding. Early robots could reliably manage only a single isolated task, such as pouring water or folding a piece of clothing, he said, whereas current models can operate across larger spaces and carry out continuous, multi step workflows. Even so, he acknowledged that significant technical hurdles remain before robots can operate reliably in the varied, cluttered conditions of a typical household.

The competing predictions reflect a broader race unfolding across China’s robotics industry, where hardware advances have outpaced the software needed to make humanoid robots genuinely useful outside controlled settings. As firms including Spirit AI, ACE Robotics and Unitree pour resources into solving the data and intelligence bottleneck, the industry’s next major test will be whether any of these timelines hold up once robots move from staged demonstrations into the messier reality of factories, shops and eventually private homes.