Humanoid robots could take three to five years to reach a level of capability comparable with GPT-3.5, according to Yao Maoqing, co-founder of Chinese robotics firm AGIBOT.
Speaking at the Fortune Leaders Forum in Macau on September 8, Yao said embodied AI – systems designed to operate in the physical world – was still developing but would eventually reach the standard associated with the model behind ChatGPT.
“As models keep maturing, [embodied AI] will reach GPT-3.5,” Yao said. “People differ somewhat on timing, but overall it’s within the 3-to-5-year window.”
GPT-3.5 was described as a turning point because it enabled an AI service to handle common, everyday tasks, rather than only specialised ones, with a success rate of between 80% and 90%.
The prediction comes as Chinese companies expand their efforts to commercialise humanoid and quadruped robots. AGIBOT shipped 9,700 units in the first six months of the year, making it the leading seller of humanoid robots, according to Counterpoint Research data released in late August. The company is considering a Hong Kong flotation.
Robots have attracted public attention through demonstrations including sprinting, weightlifting, kickboxing and dancing at the World Humanoid Robot Games in Beijing in August. But manufacturers are increasingly focused on finding practical uses beyond sporting and entertainment displays.
Shanghai-based Keenon Robotics has developed machines for hotel services. Wan Bin, the company’s chief operating officer, said one hotel in the city was using five robots to greet guests, deliver room service, clean rooms and manage the restaurant floor.
“This scene is one that will become increasingly common in the future,” Wan said.
Other companies are targeting factories and industrial workplaces. Jianxin Pang, vice-president and vice-dean of research at UBTech, said the company had invested heavily in the industrial sector while initially concentrating on common applications with sufficiently large markets to support a return on investment.
Yao acknowledged that deploying humanoid robots reliably in factories remained a significant challenge. Manufacturers, he said, judged them on success rate, cycle time, stability and cost, rather than on the technology used to build them.
“From that standpoint, we absolutely have to train our robots to 100% efficacy first,” he said.
AGIBOT said a six-day livestream in late June showed its robots completing more than 64,000 manufacturing tasks with a 99.99% success rate. Yao said achieving the result involved eight-hour sessions during the middle of the night for a month.
He remains confident that embodied AI will benefit from the same kind of rapid improvement seen in digital intelligence and large language models. As data volumes and model parameter counts increase, he said, the technology’s intelligence would undergo a significant step up.
