DexForce Is Training Humanoid Robots in Virtual Worlds to Skip the Data Bottleneck

On a screen inside DexForce's offices in Shenzhen, a virtual robot places objects on a table. It repeats the task again and again, and each repetition is slightly different: the angle of the arms changes, the color of the table changes, the objects themselves change. None of it is happening in the physical world, and that is precisely the point.

DexForce, a roughly 500-person startup based in China's southern tech hub, is betting that simulated training can solve one of the humanoid robot industry's hardest problems: teaching machines to perform a wide range of real-world tasks without human assistance. Chinese laboratories have already built humanoids that can outrun and outjump people. The harder commercial test, and the one that will decide who makes money, is whether these machines can reliably do useful work in factories, offices, and homes.

Most Chinese robot makers are attacking the problem by collecting enormous volumes of real-world data, recording human workers as they repeatedly perform tasks so robots can learn from the demonstrations. DexForce goes in the opposite direction, generating synthetic data inside virtual environments to cut costs and improve training efficiency. On October 15, reporting by Reuters from Shenzhen put the company's strategy, and its ambitions, on the record.

The 10-trillion-sample problem

Jia Kui, DexForce's chief executive, is a machine learning professor at the Chinese University of Hong Kong, Shenzhen, who founded the company in 2021. He frames the data problem in stark arithmetic. Reaching what he calls "physical artificial general intelligence," a humanoid that can take a spoken command and complete an unfamiliar task in an unfamiliar environment, would require something on the order of 10 trillion samples collected purely from the real world. Collecting demonstrations by hand for an endless list of human tasks, he argues, is fundamentally inefficient and can never get there.

The comparison he reaches for is ChatGPT. Its 2022 release pulled generative AI into the mainstream almost overnight, and Jia argues that synthetic training data could power a comparable step change in physical capability, and sooner than the field's current trajectory would deliver. Getting there, he says, requires AI systems that can perceive and navigate their surroundings and adapt to variation, exactly the skills that thousands of randomized simulation runs are designed to teach.

Ten trillion real-world samples is roughly what DexForce's CEO says the industry would need to reach true physical general intelligence, a number manual data collection can never hit.

The economics lean in his direction. A human trainer may need to repeat a single movement more than 100 times to generate precise training examples, and dozens of humanoid robot data training centers are currently being built across China at great expense, often heavily reliant on government subsidies. Most Chinese robot makers use a blend of simulated and real-world data, and many say access to high quality training data remains their key bottleneck. DexForce is simply the most aggressive bet that the blend should tilt hard toward simulation.

From virtual to real world

The company's method is a loop. DexForce develops AI software that simulates physical environments and models that predict how objects move and interact. It builds a digital version of a workplace or task, trains the robot inside that simulation across thousands of randomized variations, and then transfers what the robot learned onto physical machines.

Randomization is the key idea. By varying arm angles, surfaces, lighting, and objects in simulation, the robot learns the underlying task rather than memorizing one specific setup. When it encounters a real table it has never seen, the variation it trained on is supposed to make the new scene feel familiar. This is the standard playbook of "domain randomization," and DexForce is pushing it further than most.

The honest caveat is the sim-to-real gap, one of robotics' oldest open problems. Physics simulators are approximations, and contact forces, friction, and flexible objects are notoriously hard to model exactly. Skills that look flawless in a virtual world can degrade when a real gripper meets a real object. DexForce's wager is that enough varied simulation, combined with models that predict physical interactions accurately, closes the gap far enough for commercial work. Whether that holds at scale is the question the whole industry is watching, and the company will ultimately be judged on the performance of its physical robots, not its virtual ones.

A 500-person startup with billion-yuan backing

DexForce is not a lab curiosity. The company is backed by Lenovo Capital, the investment arm of Lenovo, and Lens Technology, alongside several state-owned funds, and it recently raised 1 billion yuan, about $147 million, in a Series B funding round at a valuation above 10 billion yuan. Bloomberg has reported that the company was considering a Hong Kong IPO; Jia declined to comment on that report when asked by Reuters.

The technology already has customers. DexForce's software is used in industrial settings by Chinese and international manufacturers including Toyota, Samsung, Panasonic, Foxconn, Haier, and Hisense. Its humanoid robots, meanwhile, are making coffee and serving customers at malls and tourist sites in more than 15 cities across China. That footprint, part factory floor and part public showcase, gives the company a steady stream of real-world feedback to check its simulations against.

What it means for the humanoid race

The timing matters because the market is accelerating. Industry figures reported this week from IDC put global humanoid robot shipments at nearly 25,000 units in the first half of 2026, up 432.1 percent from a year earlier, with market value above $740 million. IDC projects shipments surpassing 750,000 units by 2030, and says competition is shifting toward real-world deployment, scalable applications, and commercial value creation.

That shift is exactly the test DexForce has chosen. China can already build humanoids that dance, box, and sprint; the prize now is a machine that can be told to restock a shelf in a store it has never visited and simply do it. Whoever solves the training data bottleneck first will own the economics of the entire category, because every competitor faces the same 10-trillion-sample wall.

Whether synthetic data delivers the promised breakthrough is still unproven at scale. But the industry's center of gravity has clearly moved. The debate is no longer whether humanoids can walk. It is how fast they can learn, and DexForce is betting the answer will come from machines that practice in worlds that do not exist.