“As models keep maturing, [embodied AI] will reach GPT-3.5,” said Yao Maoqing, co-founder of the Chinese robotics firm AGIBOT, during the Fortune Leaders Forum in Macau on Sept. 8. “People differ somewhat on timing, but overall it’s within the 3-to-5-year window.” 

Yao’s use of “GPT-3.5” was a reference to the model underinning ChatGPT; OpenAI’s chatbot was the first time an AI service could perform common, everyday tasks—and not just specialized ones—with a success rate between 80% and 90%.

AGIBOT is part of a wave of Chinese startups developing humanoid and quadruped robots. With 9,700 units shipped during the first six months of the year, AGIBOT is the top seller of humanoid robots, according to data from Counterpoint Research released in late August. The startup is considering an IPO in Hong Kong.

While robot dance performances and boxing matches get headlines, manufacturers are frantically searching for real-world applications for their tech. Shanghai-headquartered Keenon Robotics, for instance, has rolled out robots aimed at automating hotel services.

Wan Bin, Keenon’s chief operating officer, noted that one Shanghai-based hotel is 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.”

Other robot manufacturers hope to apply their products in more industrial settings. “We’ve invested very heavily in developing robots for the industrial sector,” said Jianxin Pang, UBTech’s vice president and vice dean of research. “But first, we’re starting with relatively common scenarios with a big enough market—that helps us ensure return on our investment.” 

Yet many humanoid companies acknowledge that building robots for industry is no simple task.

“It’s genuinely difficult for today’s humanoid firms to develop a robot that can run stably inside a factory,” Yao acknowledged. “Industry players are brutally honest: They care about four metrics—success rate, cycle time, stability and cost—rather than what technology is used. From that standpoint, we absolutely have to train our robots to 100% efficacy first.”

In late June, AGIBOT ran a six-day livestream that demonstrated that its robots hit a 99.99% success rate completing over 64,000 manufacturing tasks. Reaching that milestone required eight-hour sessions in the middle of the night for a month, Yao said on-stage. 

Still, Yao was optimistic that embodied AI would follow the same scaling law that transformed large language models. 

“Most industry teams believe that embodied AI will follow the exponential scaling law we discovered in digital intelligence and large language models,” Yao said. “As data volume rises and your model’s parameter count grows, there will definitely be a step-up in intelligence.”