Spirit AI sees humanoid robot brain breakthrough by mid-2027, home use years away
Spirit AI co-founder Gao Yang says humanoid robot brains could see a GPT-3-scale breakthrough by mid-2027, but household deployment may take at least eight years.
Humanoid robot intelligence could reach a breakthrough as early as mid-2027, though widespread deployment in homes remains at least eight years away, according to Gao Yang, co-founder and chief scientist of Chinese embodied AI company Spirit AI.
Speaking at the company's Beijing offices, Gao described the robot brain as "the weakest link in the complete robotics stack." He compared the anticipated advance to OpenAI's GPT-3.0 model, which powered ChatGPT, suggesting a similar leap for robotics in roughly two years.
Spirit AI's robots currently achieve a 90% success rate on simple tasks in structured living-room environments. Gao outlined a phased path: the next one to two years will mark the initial window for industrial applications, followed by deployment in commercial service settings for simpler tasks. Homes, he said, present a far harder challenge.
The company has tens of its Moz1 wheeled humanoid robots working on production lines at battery maker CATL and retailer JD.com, which is also an investor. Founded in 2024, the 300-person startup has raised over $670 million and is valued at 20 billion yuan ($2.9 billion). Gao declined to comment on any initial public offering plans.
Gao noted rapid progress since the company's founding, when a robot could perform only a single isolated task such as pouring water or folding clothing. Today, he said, robots operate across large spatial areas and execute continuous complex workflows. Challenges remain in fine-motor actions like unscrewing a bottle cap and handling unseen tasks.
A key constraint is data. Spirit AI relies overwhelmingly on real-world data rather than virtual simulations, which many competitors use to cut training costs. Gao said simulators handle rigid bodies well but struggle with flexible objects like deformable electric cables.
To gather training data, the company employs around 1,000 contractors nationwide who use wearable data-collection equipment in households and on production lines. At its Beijing robot data training centre, dozens of young people fitted with sensors were seen repeating motions such as opening fridges, unlocking safes and cutting vegetables. While other facilities may require more than 50 repetitions to obtain one clean movement, Spirit AI found that using "dirty data" with a more diverse range of motions helped its models improve faster.