Unitree and Zhiyuan Robots Share One Brain in Unedited Demo, Report Says

The video begins with a Unitree robot cleaning a window with a squeegee. When it fails to wipe a spot clean, it leans out of the window to reach the dirt, showing a combination of precise force control, spatial awareness, and dynamic reasoning. The robot then returns the tool to its original place, completing a full task loop.

In another scene, a Zhiyuan robot takes clean items from a washing machine and puts them away in cabinets and on a sofa. When a phone alarm rings, both robots pause their current tasks to handle a separate tidying job, then resume exactly where they left off—demonstrating the ability to handle interruptions and multi-task switching.

The most striking moment comes when the two robots cooperate. The Unitree robot, unable to hold more objects, receives help from the Zhiyuan robot, which picks up a scarf from the table and hangs it around Unitree's neck. Later, when Unitree tries to place the scarf on a high shelf, it finds a box to stand on, tries to lift it, fails due to its cannot-bend body, and ultimately kicks the box closer—an apparent example of tool use and self-exploration.

The Zhiyuan robot also seems to learn to hang a towel on its own shoulder to save effort, and picks new slippers by their hanger string rather than the shoe body. These behaviors, according to the report, suggest the model can discover efficient strategies through interaction with the environment.

According to people familiar with the matter, the model behind the video breaks away from mainstream VLA, WAM, and traditional world models. It reportedly uses physical-constraint-aware dynamic learning, cross-embodiment unified modeling, and a long-horizon robust closed-loop framework. The model was trained on only a few tens of hours of video data, raising questions about the scaling law in embodied AI.

The report notes that the video may be the first publicly known example of a single model operating two different robot bodies in a cooperative task. While the team remains unnamed, the implications for the industry are significant, as it suggests a path toward decoupling AI models from specific hardware.