By Laurie Chen
BEIJING, Sept 18 (Reuters) - Humanoid robots will be able to complete most general-purpose tasks from verbal instructions as soon as next year, although becoming useful around the house will take much longer, said the head of a leading Chinese company developing robot brains.
After impressive hardware advances allowing Chinese humanoids to sprint, dance and do backflips on command, robot firms are increasingly focusing on the software that determines robots' intelligence and enables
them to do more economically productive tasks, a field known as "embodied AI".
"The brain is indeed the weakest linkv in the complete robotics stack," Gao Yang, co-founder and chief scientist of Spirit AI, told Reuters at its Beijing offices on Thursday.
The robotics industry is searching for its "ChatGPT moment": a software breakthrough that could turn frontier technology into products useful to a much wider range of customers. OpenAI's launch of ChatGPT in 2022 brought generative AI to a mass audience and demonstrated its commercial potential.
"We anticipate reaching the GPT-3.0 milestone by mid-2027. You will be able to speak to a robot in natural language, and it will execute a series of reasonable physical actions to attempt the task," said Gao.
The company employs around 1,000 contractors nationwide who wear data-collection equipment at home or in factories, providing data on how humans move about their environments to help train robots.
During a visit to a data training centre in its Beijing office, dozens of young people fitted with sensors were repeating motions such as opening fridges, unlocking safes, and cutting vegetables with knives.
Spirit AI's robots have achieved a 90% success rate for simple tasks in structured living-room environments.
"The next one to two years mark the initial window for industrial applications. Two years from now, we'll see robots deployed in commercial service settings doing simpler tasks. Entering homes is far harder than both," said Gao, who is also an assistant professor of robotics at Tsinghua University.
FOCUSED ON BRAINS
Spirit AI currently has tens of its own Moz1 wheeled humanoid robots deployed on production lines at battery maker CATL and retailer JD.com, which is also an investor.
The 300-person startup has raised over $670 million since its 2024 founding, making it one of China's most rapidly capitalised embodied intelligence firms. It is currently valued at 20 billion yuan ($2.9 billion). Gao declined to comment on any plans for an initial public offering.
"Progress is extremely fast. When Spirit AI was founded, a robot could perform only one isolated task well, like pouring water or folding a piece of clothing," said Gao.
"Today, robots operate across large spatial areas and execute continuous complex workflows."
Difficulties remain in perfecting fine-motor actions such as unscrewing a bottle cap, and dealing with unseen tasks, he said.
Spirit AI overwhelmingly relies on real-world data to train its robot brains instead of virtual simulations. Many competitors use simulators to reduce model training costs, but Gao said some tasks can only be learned by studying actual interactions with physical objects.
"Simulators handle rigid bodies well, but flexible objects like deformable electric cables remain a problem," said Gao.
At other robot-training facilities in China, operators may need to repeat a movement more than 50 times to get one "clean" movement with the required precision. Spirit AI found that using "dirty data" with a more diverse range of motions enabled its models to improve faster, Gao said.
ROGUE ROBOTS?
AI companies in the United States have called for a slowdown in development of advanced language models in recent days after hacking incidents showed the risk of uncontrolled agents.
Gao said the risk of an AI going rogue was less of a pressing concern for robots in the physical world for now, as the software models are currently too immature.
But safety for embodied AI is becoming a bigger issue as robots will eventually interact with more humans in crowded and unpredictable settings such as commercial and home environments.
"Our robots feature whole-body force control. If the robot encounters excessive interaction force with the environment, emergency braking triggers automatically as a baseline safety policy," said Gao.
"Once foundation models reach a mature, autonomous 'GPT-4.0' era, researching advanced AI safety and alignment will become much more actionable."
(Reporting by Laurie ChenEditing by Eduardo Baptista, Christian Schmollinger, Peter Graff)
