By Laurie Chen
SHENZHEN, Oct 15 (Reuters) - At the offices of Chinese startup DexForce, a screen shows a virtual robot placing objects on a table. It repeats the task as the angle of its arms, the colour of the table and the objects themselves change.
DexForce, based in the southern tech hub of Shenzhen, is betting that such simulated training can help solve one of the humanoid-robot industry's biggest challenges: teaching machines to perform a wide range of real-world tasks without human assistance.
China has developed humanoids capable of outperforming humans in running and jumping but the harder commercial test lies in whether they can reliably carry out useful work in factories, offices and the home.
Many Chinese robot makers are trying to meet that challenge by collecting large volumes of data from human workers repeatedly performing tasks in the real world. DexForce is relying on synthetic data generated in virtual environments to cut costs and improve training efficiency.
Chief Executive Jia Kui told Reuters the approach could more quickly power a breakthrough in robot capability, akin to the one ChatGPT triggered after its 2022 release helped propel generative AI into the mainstream.
"Trying to achieve 'physical artificial general intelligence' purely through real-world data collection would require 10 trillion samples ... manual (data) collection for infinite human tasks is fundamentally inefficient," said Jia, referring to a technological milestone where a humanoid can complete new tasks in an unfamiliar environment on command.
Reaching that point will require AI systems that can perceive and navigate their surroundings and adapt to variation.
FROM VIRTUAL TO REAL WORLD
DexForce develops AI software designed to simulate physical environments and models intended to predict how objects move and interact.
It creates a digital version of a workplace or task, trains a robot in the simulation and then transfers that learning to physical machines.
The company says the process is faster and cheaper than collecting data from human trainers, who may need to repeat a movement more than 100 times to generate precise examples.
Dozens of humanoid robot data training centres are currently being built nationwide at great expense, and are often heavily reliant on government subsidies.
Most Chinese robot makers use a combination of simulated and real-world data, and many say access to high-quality training data remains a key bottleneck.
STARTUP BACKED BY CHINESE CONGLOMERATES
Jia, a machine-learning professor at the Chinese University of Hong Kong, Shenzhen, founded DexForce in 2021.
The roughly 500-employee startup is backed by Lenovo Capital, the investment arm of Lenovo, and Lens Technology, as well as several state-owned funds and raised 1 billion yuan ($147 million) in a recent Series B funding round at a valuation of more than 10 billion yuan.
Bloomberg reported it was considering a Hong Kong IPO, but Jia declined to comment.
The firm's software is used in industrial settings by Chinese and foreign manufacturers including Toyota, Samsung, Panasonic, Foxconn, Haier, and Hisense, while its humanoid robots make coffee and serve customers at malls and tourist sites in more than 15 cities nationwide.
(Reporting by Laurie Chen in Shenzhen; Editing by Eduardo Baptista and Kate Mayberry)
