DexForce bets virtual worlds can break humanoid robots’ data bottleneck

The story

Humanoid robots can now walk, balance and perform carefully staged demonstrations. The harder problem is teaching them enough useful work to justify a place on a factory floor. Shenzhen-based DexForce is attacking that bottleneck with virtual data: it builds simulated versions of scenes and tasks, trains policies there, then transfers the result to physical robots. Reuters reported on October 15 that the company sees this route as a way around the immense cost and time required to collect demonstrations entirely in the real world.

The logic is straightforward. A robot learning a new manipulation skill needs examples of what to do, observations of how objects respond and feedback when an attempt fails. Human teleoperation can provide high-quality demonstrations, but recording and labelling them does not scale cheaply across thousands of objects, layouts and edge cases. Simulation can multiply a smaller set of real observations by changing camera positions, object geometry, lighting and starting conditions without occupying a physical workcell or risking hardware.

DexForce calls its broader pipeline Real2Sim2Real. According to the company, its DexVerse platform reconstructs real environments in simulation, while EmbodiChain generates and manages synthetic data before policies return to machines such as its W1 Pro humanoid. Its website lists 40 degrees of freedom, a 1,000-hertz joint-control frequency and force-control resolution of 0.1 newton for the latest W1 Pro generation. Those are vendor specifications, not independently verified measures of performance in customer operations.

A newly disclosed research system supplies the most concrete technical account of the strategy. UNITAS, a 1.7-billion-parameter model described in an arXiv preprint submitted on October 8, is designed to predict both an action and how a 3D scene will change after that action. The DexForce-led authors call it a 3D-native world action model: instead of translating among separate representations for language, video and robot controls, it places manipulation in a shared metric 3D frame.

UNITAS separates that representation into two linked flows. “Action flow” describes 3D trajectories for human hands or robot grippers; “scene flow” describes how points in the surrounding scene move when conditioned on those trajectories. In principle, the model can learn from human activity, robot demonstrations and simulated interaction while preserving spatial relationships. That matters for tasks such as grasping, folding or placing, where a small error in contact position can be the difference between success and a dropped object.

The paper reports strong results, including as much as 49% lower displacement error than PointWorld on the RoboTwin benchmark, a 99.8% success rate on LIBERO and an 85% average across the authors’ real-world tasks. These numbers require restraint. The paper is marked under review, the comparisons and real-world evaluations come from the system’s own developers, and broad independent replication is not yet available. A benchmark win also does not establish economical, unattended operation through a full industrial shift.

Simulation creates its own risks. Contact dynamics, friction, transparent or reflective materials, sensor latency, mechanical wear and inconsistent lighting can all expose the “reality gap” between an idealised environment and a factory. Synthetic data is therefore not a substitute for real measurements. Its value depends on how faithfully the virtual environment is calibrated and how quickly failures from deployed robots flow back into the next training cycle.

DexForce nevertheless has resources to test the model at scale. Reuters reports that the company, founded in 2021 and led by robotics researcher Kui Jia, has roughly 500 employees and recently raised about 1 billion yuan, or $147 million, in a Series B round at a valuation above 10 billion yuan. Reuters also says DexForce software is being used in industrial settings by companies including Toyota, Samsung, Panasonic, Foxconn, Haier and Hisense, and that its robots serve customers in more than 15 Chinese cities. Those deployments provide useful context, but they do not by themselves disclose task-level uptime, intervention rates or commercial margins.

INNOVOX analysis: the decisive race in humanoid robotics may be less about who unveils the most agile machine and more about who builds the most credible learning loop. Synthetic data can turn rare demonstrations into many training variations; a 3D world-action model can make those variations more physically meaningful; and deployed robots can return the failures needed to refine both. If the loop works across different bodies and workspaces, software and data infrastructure could become more defensible than any single robot design.

What to watch next is reproducibility and operating evidence. UNITAS code, checkpoints and evaluation protocols would allow outside teams to test the claimed gains. Cross-robot transfer would show whether the representation is genuinely general rather than tuned to one embodiment. In factories, the telling figures will be interventions per hour, task-cycle consistency, recovery from unfamiliar objects and the cost of adding a new skill. Until those measures are public, DexForce’s virtual-data strategy is a technically plausible attempt to solve the right problem—not proof that general-purpose humanoid labour has arrived.

INNOVOX analysis

Humanoid hardware is converging faster than dependable training data. DexForce’s strategy is significant because it treats simulation not merely as a test bed but as a data factory and a place to learn how actions change a 3D scene. If that loop transfers across factories and robot bodies, the competitive advantage shifts from spectacular demonstrations to the quality of simulation, calibration and failure feedback. For now, that remains a hypothesis supported mainly by the company’s own evidence.

What to watch

Watch for independent reproduction of UNITAS, public code and checkpoints, cross-robot transfer results, and production metrics such as intervention rate, task-cycle time and performance under unfamiliar lighting, materials and layouts. These will show whether synthetic-data gains survive contact with real factories.