Humanoids will have their ‘GPT-3.5’ moment within 5 years, says AGIBOT co-founder Yao Maoqing

Humanoid robots recently stole the spotlight at the World Humanoid Robot Games held in Beijing, where they showcased their athletic abilities through tasks like sprinting, weightlifting, and kickboxing. While these displays were impressive, one Chinese robotics entrepreneur remains cautious about the timeline of widespread public acceptance.

According to Yao Maoqing, co-founder of AGIBOT, a leading Chinese robotics company, AI-powered robots will reach the same level of sophistication as the GPT-3.5 model within the next three to five years. This statement came during a recent Fortune Leaders Forum in Macau, where Yao utilized "GPT-3.5" as a benchmark, drawing a parallel to OpenAI's ChatGPT, which demonstrated an 80% to 90% success rate in performing common, everyday tasks.

AGIBOT is currently one of the top sellers of humanoid robots, having shipped 9,700 units within the first half of the year, according to Counterpoint Research. The company is actively seeking an IPO in Hong Kong. Despite the focus on robot performances and competitions, manufacturers are actively seeking practical applications for their technology.

For instance, Keenon Robotics, based in Shanghai, has developed robots designed to automate hotel services, including greeting guests, delivering room service, cleaning rooms, and managing the restaurant floor. WAN Bin, the chief operating officer of Keenon Robotics, noted that one Shanghai-based hotel employed five robots to handle various guest-related tasks.

This growing trend of robot deployment in increasingly common scenarios is expected to become more prevalent in the future. However, several humanoid robot manufacturers acknowledge the challenges in developing robots for industrial applications. Yao Maoqing admitted that creating a stable robot for factory environments is a significant challenge.

He emphasized that industry players are primarily concerned with four key metrics: success rate, cycle time, stability, and cost. To achieve success, AGIBOT conducted a six-day livestream demonstrating its robots achieving a 99.99% success rate in completing over 64,000 manufacturing tasks. This feat required an eight-hour session over a month, conducted during the early hours of the night.

Nevertheless, Yao remains optimistic about the future of embodied AI, comparing its scaling law to the exponential growth observed in large language models. He believes that as data volume increases and model parameters grow, embodied AI will demonstrate significant intelligence advancements.