An industry breakthrough ⁠comparable to OpenAI’s landmark GPT-3.0 model could come around mid-2027

BEIJING: A major breakthrough in humanoid robot intelligence could come as early as mid-2027, but widespread use of humanoid robots in homes may still be at least eight years away, according to Chinese embodied AI company Spirit AI.

The company says the biggest challenge is no longer the physical abilities of humanoid robots but the software, or “brain,” that allows them to understand their surroundings and perform useful tasks.

“The brain is indeed the weakest link in the complete robotics stack,” Gao Yang, co-founder and chief scientist of Spirit AI, told Reuters in Beijing.

Gao said a breakthrough in robot intelligence comparable to the impact of OpenAI’s GPT-3 could emerge around mid-2027.

Spirit AI’s robots currently have a 90% success rate when performing simple tasks in structured living-room environments, according to the company.

Commercial robots could arrive sooner

Gao said the next one to two years could mark the beginning of wider industrial applications for humanoid robots.

He expects robots to be used in commercial settings for relatively simple tasks within about two years, while household applications will require significantly more development.

“Entering homes is far harder than both,” Gao said. He is also an assistant professor of robotics at Tsinghua University.

Focus shifts to robot intelligence

Chinese robotics companies have made rapid progress in developing humanoid machines capable of running, dancing and performing backflips. However, the industry is increasingly focusing on developing AI systems that can make robots more capable and productive in real-world environments.

Spirit AI currently has dozens of its Moz1 wheeled humanoid robots operating on production lines at battery manufacturer CATL and retailer JD.com, which is also an investor in the company.

The 300-person startup has raised more than $670 million since it was founded in 2024 and is currently valued at around 20 billion yuan ($2.9 billion), according to the report.

Gao said robots have progressed significantly since the company’s launch. While earlier systems could perform individual tasks such as pouring water or folding clothes, newer models can operate over larger areas and complete longer, more complex workflows.

However, robots still struggle with fine-motor tasks, such as unscrewing bottle caps, and with situations they have not encountered during training.

Real-world data remains a challenge

Spirit AI mainly uses real-world data to train its robot models rather than relying heavily on virtual simulations.

Gao said simulators can handle rigid objects effectively, but flexible objects such as electric cables remain difficult to reproduce accurately in virtual environments.

To collect more real-world data, Spirit AI employs around 1,000 contractors across China who use wearable equipment to record human movements in homes and production facilities.

At the company’s Beijing office, Reuters observed a robot training centre where people wearing sensors repeatedly performed tasks such as opening refrigerators, unlocking safes and cutting vegetables.

Some robot-training facilities require operators to repeat the same movement more than 50 times to produce one precise training example.

However, Gao said Spirit AI has found that using more varied “dirty data” — rather than only highly precise movements — can help its AI models improve more quickly.

The company believes solving the data challenge will be key to developing robots capable of safely and reliably carrying out a much wider range of tasks in real-world environments.