DexForce bets on simulated data while China subsidises real-world robot data centres

DexForce, a Shenzhen physical AI company, trains its humanoid robots mainly on synthetic data generated in virtual environments rather than on recorded human demonstrations, chief executive Jia Kui told Reuters in a feature published on October 15, 2026. Jia said that reaching “physical artificial general intelligence” purely through real-world data collection “would require 10 trillion samples”.

Why it matters: China is spending heavily on the opposite approach. Reuters reports that dozens of humanoid data-training centres are being built across the country, at great expense and with heavy reliance on government subsidies.

How does DexForce train its robots?

DexForce builds a digital version of a workplace or task, trains the robot there, then transfers what it learned to physical machines, according to Reuters. The company says this is faster and cheaper than collecting data from human trainers, who may need to repeat a movement more than 100 times to produce precise examples. Jia called manual collection for “infinite human tasks” “fundamentally inefficient” and compared the breakthrough he expects from synthetic training to the one ChatGPT set off in 2022.

The core engine is DexWorldModel, a world model that Gasgoo reported had topped WorldArena, a benchmark for embodied world models. Gasgoo reported that the Series B money would go to the company’s world-model algorithms, its physics engine, data infrastructure and real-world humanoid deployment. Reuters lists Toyota, Samsung, Panasonic, Foxconn, Haier and Hisense among the manufacturers using its industrial software. Bloomberg has reported that DexForce is considering a Hong Kong IPO; Jia declined to comment.

How is that different from China’s data-factory model?

The state-backed approach pays workers to perform tasks while robots and sensors record them. Bloomberg counted 64 such centres in July, with about 20 more under construction, as we reported in China’s answer to the robot data problem. Japan’s J-HRTI opened a similar teleoperation data factory in August. The logic is that high-quality training data is the bottleneck, a view many Chinese robot makers share, according to Reuters, and that real demonstrations are the surest way to get it.

The line is less sharp in practice. Most Chinese robot makers already mix simulated and real data, Reuters notes, and DexForce did not say how much real-world data its own models use. No independent results for its simulation-trained robots have been published. The checkable numbers are commercial: whether it reaches its 300 million yuan revenue target for 2026, and whether its sim-to-real transfer holds up as its robots move from coffee counters to the sorting, assembly and dual-arm tasks Gasgoo says they are being prepared for.