A “ChatGPT moment” for China’s humanoid robots – the tipping point at which the technology becomes widely usable – remains years away as persistent challenges in adapting to new tasks and training efficiency continue to hold back the industry, leading experts said on Wednesday at the Boao Forum for Asia in Hainan.
Despite rapid advances in recent years, humanoid robots were still far from large-scale deployment, with both hardware and software limitations yet to be fully resolved, panellists said during a discussion on the sector’s future.
“The core issue is that robotics data is extremely high-dimensional, while text data [used to train large language models] is essentially one-dimensional,” said Shao Hao, chief scientist at the robotics lab of Chinese smartphone maker Vivo.
“Looking back, deep learning began gaining momentum around 2012, but the breakthrough moment didn’t arrive until around 2019. The key difference maker was data.”
In the robotics industry, references to OpenAI’s “ChatGPT” have become shorthand for the point at which a technology overcomes key technical bottlenecks and achieves mass adoption.
By massively expanding the volume of training data – including large amounts of human-labelled inputs – OpenAI developed models capable of generalising across previously unseen tasks, underpinning ChatGPT’s launch in late 2022.
However, while OpenAI was able to draw on vast, readily available online data at scale, the robotics industry has yet to identify a similarly cheap and reliable source of its own.
Wang Xiaogang, co-founder of SenseTime and chairman of its robotics spin-off Ace Robotics, said the industry had so far accumulated only hundreds of thousands of hours of training data, largely derived from human teleoperation.
This paled in comparison to adjacent sectors such as autonomous driving, which generated millions of hours of data on a daily basis using cutting-edge simulations, he said.
“A ‘ChatGPT moment’ for robots will only happen when we move beyond manual data collection methods,” Shao said, estimating that it could take about a decade.
Wang, however, struck a more optimistic note, suggesting such a breakthrough could come within two years, as experimental data-generation techniques expanded available data sets to tens of millions of hours by 2027.
These approaches include training robots to observe and replicate human behaviour across diverse real-world scenarios – a shift that is placing greater emphasis on developing more capable “brains”, or software systems, that can adapt to unfamiliar environments.
“[The goal is] a robotic brain that can operate across different hardware bodies, making behaviour much more humanlike,” Wang said, adding that future systems may be able to “self-evolve” as their underlying models improve.
China remains at the forefront of humanoid development, with leading players such as Unitree Robotics and AgiBot. Morgan Stanley forecasts that the country’s humanoid robot sales will more than double this year to 28,000 units.
Also on Wednesday’s panel, former New Zealand prime minister Jenny Shipley urged industry participants to carefully consider the societal implications of humanoid robots, including how female forms were represented. – South China Morning Post
