DexForce trains humanoids with virtual data to cut costs, Reuters reports

Reuters reported on October 15, 2026, that Chinese startup DexForce is using virtual data to train humanoid robots and reduce costs. The report said the approach is intended to reduce the burden compared with training centered on physical robots. Details in the article on the scale of training and evaluation conditions were limited.

Training with virtual data could affect how much the number of physical tests and the burden of data collection required for humanoid development can be reduced. Reuters’ report that DexForce is adopting this approach raises the possibility that development’s focus could shift from the robots themselves to the training infrastructure.

What is new in this report is the claim that DexForce is using virtual data to train humanoids and lower costs. This can be read as a shift away from training that relies heavily on physical robots toward a greater emphasis on data generation and learning. However, the Reuters excerpt does not say how much data is being used or how much physical-robot training has been replaced, making it difficult to assess the scale of the benefits so far. In a related article, DexForce unveils 3D world model “UNITAS”, we reported that DexForce had introduced a 3D world model for robots. The virtual-data training described in this report is closer to the practical application of such a world model and appears consistent with a concept that integrates perception, action and learning. However, implementation details and evaluation conditions were limited when UNITAS was introduced, so the key question here is likewise whether the approach is still at the conceptual stage or has reached practical deployment. From an industry perspective, the report suggests that the bottlenecks in robotics development lie not only in robot design but also in securing training data and the cost of generating it. Training with virtual data could reduce the burden of trial and error in the field and damage to physical robots. In turn, the fidelity of simulations, transfer to real-world environments and reproducibility of failure cases become key issues. The next point to clarify is which processes DexForce is virtualizing and what it is verifying with physical robots. Unresolved questions include how the virtual data is generated, its ratio to real-world data used for training, the target robot models, performance evaluation criteria and the specific extent of cost reductions. Without this information, the company’s approach can be viewed as an effort to improve training efficiency, but its direct relevance to mass production or commercial deployment is difficult to assess.

Japanese robotics developers could also make greater use of training in virtual environments and 3D world models, but this report alone provides limited insight into which processes DexForce’s approach affects. For Japanese companies, the issues will include how to establish safety assessments that support physical testing and validation processes that bridge the gap between simulations and reality.