A Chinese artificial intelligence company is pursuing a different path to developing capable humanoid robots, using computer-generated training environments to reduce the time and expense associated with teaching machines to perform physical tasks.

DexForce, a robotics startup headquartered in Shenzhen, creates virtual settings where robots can repeatedly practice activities under changing conditions. In demonstrations at its facilities, simulated machines sort and reposition objects while details such as their surroundings, colors and movements vary.

The strategy addresses a central problem for the robotics industry. While humanoid machines have demonstrated increasingly impressive physical abilities, teaching them to perform dependable, useful work outside controlled demonstrations remains a much harder challenge.

The company’s approach to robot training

Many robotics developers collect data by having people demonstrate tasks in actual workplaces. DexForce is placing greater emphasis on synthetic data generated through digital simulations, allowing machines to encounter numerous variations without requiring a person to physically repeat every action.

Chief Executive Jia Kui argues that collecting enough real-world examples to create broadly capable robots would be prohibitively demanding.

“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,”

Jia told Reuters, describing a future in which robots could follow instructions to complete tasks they had never encountered before.

Reaching that goal would require machines to interpret their surroundings, navigate different environments and adjust their behavior when conditions change. Jia believes advances in training methods could help produce a significant improvement in robotic intelligence, comparable in impact to ChatGPT’s emergence in 2022 and the subsequent expansion of generative AI.

DexForce builds digital representations of physical environments and uses models to predict how objects behave and interact. After a robot develops a skill in a simulated setting, the company transfers the learned behavior to a physical machine.

According to the startup, this process can reduce the cost of producing training data. Human demonstrations may need to be repeated more than 100 times to capture movements with sufficient precision.

The broader Chinese robotics sector is also investing heavily in training infrastructure. Numerous facilities designed to collect robot-training data are under development, with government subsidies helping support some projects. Despite widespread use of both simulated and real-world training, obtaining high-quality data remains a major obstacle for manufacturers.

Investment and commercial applications

Jia founded DexForce in 2021 while serving as a machine-learning professor at the Chinese University of Hong Kong, Shenzhen. The business has grown to roughly 500 employees and received backing from Lenovo Capital, Lens Technology and several state-owned investment funds.

Its latest Series B funding round brought in 1 billion yuan, approximately $147 million, at a reported valuation exceeding 10 billion yuan.

Bloomberg has reported that the company was considering a Hong Kong stock-market listing. Jia declined to discuss the possibility.

DexForce is already targeting industrial and commercial customers rather than limiting its work to laboratory research. Its software is used by manufacturers including Toyota, Samsung, Panasonic, Foxconn, Haier and Hisense.

The company is also developing humanoid robots for service roles. Its machines have been used to prepare coffee and serve customers at shopping centers and tourist destinations in more than 15 Chinese cities.

Whether virtual training can consistently produce robots capable of handling unpredictable real-world situations remains an important test. For DexForce, the commercial opportunity lies in making robots adaptable enough for practical deployment without the enormous data-collection costs that could otherwise slow progress.