Mr. Wang Xingxing said the company's humanoid robot running OpenAI's GPT-6 model can perform meticulous tasks such as painting.

Speaking at the China International Import Expo (CIIE) in Shanghai on November 5, Mr. Wang Xingxing, founder and CEO of Unitree, said that the humanoid robot running GPT-6 is now capable of performing complex sequences of operations when painting and playing chess without needing any prior training.

For example, with painting, the robot can automatically change brush water and mix colors, entirely based on real-time model inference. "The precision needed to mix the right color without making mistakes is quite high," Xingxing said. "Seeing a robot perform at this level without any training is truly impressive."

However, instead of actual performances on stage, Xingxing only shows videos of the Unitree robot painting. According to SCMP, this may be somewhat less convincing, but still highlights how China's leading robotics companies use advanced AI models to accelerate automation on physical environments.

At an event in Shenzhen last month, Xingxing also said Unitree was testing GPT-6 and "obtaining positive results", and suggested that GPT-7, if launched next year, could give the humanoid robotics industry its own "ChatGPT moment". "The moment will come when robots can complete 80% of tasks in 80% of unfamiliar environments with only voice instructions," said the Unitree founder.

According to Xingxing, converting AI commands into precise physical actions is still "an uphill battle". At CIIE, he warned the industry's biggest obstacle is no longer digital reasoning or planning, but closing the mechanical gap "in the last millimeter".

"The fundamental problem today is the missynchronization between the input and output of the AI ​​model," he explained, predicting the industry could fix this physical "bottleneck" within a year or two.

This optimistic view is shared by company leaders in the field of humanoid robots. In particular, Wang Qian, founder of X Square Robot, said that Embodied AI is on the rise, predicting that the "GPT moment" in this field is very likely to happen within the next year.

"We want the robot to move from being controlled by code to being controlled by conventional logic," Qian told Reuters.

Longtime robot manufacturers also increased investment in Embodied AI. UBTech is currently recruiting for the position of chief scientist in charge of developing AI models with compensation up to 124 million yuan (18 million USD), including shares and benefits. Meanwhile, Unitree is arranging about 30 researchers to develop AI capable of predicting object behavior.

Chip manufacturers are more cautious, pointing out limitations in the robot's software platform. Wang Dong, co-founder of graphics chip designer Moore Threads, said that AI applications in physical environments have not developed as expected due to serious data shortages and high hardware costs. Additionally, developers still have difficulty collecting data from real-life interactions, forcing them to rely heavily on simulation data to train AI.

Robotics consultant Georg Stieler says humanoid robots have now had a "breakthrough in demonstrations". However, this does not mean they are ready for widespread commercial use.

"Given the possibility of the model handling many new tasks for the first time, if it happens during demonstrations, the optimists may be right," Stieler told Reuters. "But for customers to truly trust and use this technology in practice, it will take years instead of months."

Bao Lam summary

• American Professor: Humanoid robots are being hyped
• Unitree demonstrates self-fighting humanoid robot
• Humanoid robot Unitree falls 'sparking' while practicing running
• Superhero robot 'jumps 2 meters high, runs faster than Usain Bolt'