Zhongke Diwuji Unveils Embodied Foundation Model Using Motion Optical Flow, Maintains Top Spot at WorldArena

leaderobot.com reported on Sept. 21, 2026, that Zhongke Diwuji unveiled an embodied foundation model using motion optical flow and maintained its top spot at WorldArena. We have not been able to verify the primary source, so please refer to the original report for details.

Zhongke Diwuji’s announcement highlights that what matters in embodied AI is not only “what it understands,” but also “how it represents motion and turns it into learning.” If a representation such as motion optical flow can be used across different robot bodies, it could reduce the burden of recreating data for each robot.

The new element in this report is the representation of embodied robot motion not as a sequence of XYZ coordinates, but as a combination of motion optical flow and heat maps. This is not simply a model update; it is an approach that reconfigures the representation layer between input and output, suggesting that the company sees the difficulty of reusing existing body-dependent data as a technical challenge. Meanwhile, ranking first in WorldArena 1.0 and 2.0 indicates that the model achieved a certain result in an external environment with evaluation criteria, although the original report does not provide enough detail to fully assess the evaluation conditions. In Unitree Unveils G1+ and New Foundation Model, a related article from this newspaper, the focus was on an update to the foundation model for general-purpose humanoid robots. Zhongke Diwuji’s report, by contrast, foregrounds the design of the motion representation downstream. In other words, even when both are called “foundation models,” the key challenge differs between the layers for understanding and planning and the layer that translates actions into movements of the physical body. Taken together with Morgan Stanley Raises China Humanoid Robot Shipment Forecast to 50,000, Emphasizes Capability Gap, the reports point to a competitive landscape in which data reuse across multiple robot bodies and training efficiency are becoming important alongside shipment volumes. In terms of industry structure, whether a company can develop a representation system that accommodates differences among robot bodies is likely to affect the value of its collected data. Methods that depend heavily on coordinate points are more likely to require retraining whenever the robot body changes, while using motion optical flow as an intermediate representation could open up more opportunities to combine data from different robot bodies. However, it is not possible to assess how well this approach will work in real-world operations without checking WorldArena’s evaluation setup, success rates on physical robots, transfer performance and the composition of the training data. Please refer to the original report for details.

For Japanese companies, this is an indication that the design of motion representations and training data, not just robot bodies, will be a competitive factor. In particular, issues are likely to include how to standardize physical-robot data across multiple models and which processes can be shared.