A humanoid robot trained for a bottle-relay task using virtual demonstrations completed it with a person in seven of ten physical trials, researchers report in a September 22 preprint. The experiment tests whether coordinated actions learned from two simulated humanoids can carry over to hardware. It does not yet show two physical robots working together.
Collecting examples of two robots cooperating is difficult: both machines, their operators and the objects they handle must stay in step. The researchers built MATE, a shared simulation in which separate operators control virtual humanoids while the system records their actions and the changing positions of objects on one clock. A handoff or a delay by one robot therefore changes what the other robot sees and can do.
The team recorded 903 human-operated joint episodes across five tasks, then made 1,597 validated versions by changing object positions or timing while preserving the task sequence. The resulting dataset contains 2,500 episodes and 24.1 hours of synchronized behavior. The training method gives extra weight to moments such as grasping, releasing and responding to a partner instead of repeatedly sampling long stretches of waiting. In a simulated bottle-relay test, a robot controller trained this way finished nine of ten trials, compared with four of ten when trained on uniformly sampled moments under the same model and training conditions.
For the physical test, the researchers adapted the virtual scene's cameras and appearance to the lab, but used no real-world demonstrations or fine-tuning. A person placed a bottle at a changing intermediate position; the robot then approached it, picked it up, carried it and set it down at the target. It succeeded in seven of ten real trials, versus nine of ten in simulation.
The result suggests that virtual partner interactions can teach one robot a sequence useful outside simulation. The physical test covered one humanoid, one human partner and a single relay task, with only ten attempts. The authors identify perception, object shape and contact behavior as remaining sources of failure.
Sources
Yu and colleagues, “MATE: Multi-Agent Virtual Teleoperation Platform for Humanoid Collaboration Data Collection” (arXiv preprint, September 22, 2026); full paper and supplementary methods; research project page. Image: RuinDig/Yuki Uchida, Unitree display at Japan Mobility Show 2025, CC BY 4.0, Wikimedia Commons. This is a contextual photo, not the robot or experiment in the study.
