At DexForce’s Shenzhen office, a virtual robot repeats a tabletop task while the arm angle, object and surface change. The point is not to make a better animation. It is to create many controlled experiences that would be slow and expensive to collect from people repeatedly moving physical robots.[1]
The company calls its pipeline Real2Sim2Real. A real task supplies the starting geometry and objective; simulation produces variations and training episodes; the resulting policy is transferred back to a machine. Reuters reported that DexForce has raised 1 billion yuan—about $147 million—in Series B funding at a valuation above 10 billion yuan.[1][2]
Why simulation is attractive
What each data source contributes
| Training source | Strength | Main weakness |
|---|---|---|
| Human demonstrations in the real world | Contain real contact, mess and operating constraints | Expensive to repeat and difficult to scale across every variationSource 1 |
| Physics simulation | Can generate large numbers of controlled variations safely and quickly | The simulator cannot reproduce every material, sensor error or unexpected obstructionSource 1Source 2 |
| Factory feedback | Reveals where a transferred policy fails and supplies new cases for retraining | Requires live hardware, supervision and a process for recovering from mistakesSource 4 |
The synthetic route is especially useful for rare or awkward situations. A virtual trainer can rotate a component, change the lighting, move the target or perturb the robot without risking a production line. It can also record exact states and forces that would be hard to label by hand.[1]
The hard part is crossing the simulation gap
A simulation is an approximation. Real cables flex, packaging reflects light, grippers wear and objects arrive at angles that were not anticipated. A policy can look robust in a clean virtual world and fail when friction or camera noise shifts slightly. The practical loop therefore needs domain randomisation, real-world testing and a way to feed failures back into the simulator.[1][4]
DexForce says its W1 Pro humanoid has 40 degrees of freedom, a 1,000-hertz control loop, force sensitivity down to 0.1 newtons and vision positioning within one millimetre. Those specifications describe the company’s hardware claim; they do not by themselves establish task success over a full shift or across another factory.[2]
A factory task is more revealing than a stage demonstration
Reuters previously reported a phone-packing deployment at a Lens Technology factory, where the robot had to handle devices that did not always arrive perfectly aligned. That kind of bounded task is a useful commercial test because success can be measured: correct placement, cycle time, intervention rate, damage and performance after hours of operation.[4]
DexForce says its technology is used by manufacturers including Toyota, Samsung, Panasonic, Foxconn and Haier, and that robots using its systems operate in more than 15 Chinese cities. Those deployment counts are company-supplied. The stronger future evidence would disclose the task, operating hours, human interventions and comparison with the previous process.[1][3]
Sources & further reading
- Chinese startup DexForce trains humanoid robots using virtual data to slash costs Reuters · 15 Oct 2026 · Accessed 15 Oct 2026
- DexForce W1 Pro and Real2Sim2Real technology DexForce · Accessed 15 Oct 2026
- About DexForce DexForce · Accessed 15 Oct 2026
- China robot makers seek a path from demonstrations to mass adoption Reuters · 19 Aug 2026 · Accessed 15 Oct 2026