Humanoid Badminton demonstrates multi-skill rallies

The important advance is skill composition under sparse demonstrations, not merely teaching a robot one impressive swing.

• –Task-randomized motion augmentation creates a continuous latent skill space from limited hitting events, while a high-level planner selects skills according to shuttle state.
• –Real-world multi-skill rallies push beyond earlier humanoid badminton work focused primarily on whole-body striking and trajectory prediction. [Prior work](https://arxiv.org/abs/2511.11218)
• –The approach could generalize to other fast, contact-rich robotic tasks where demonstrations are scarce and timing is unforgiving.
• –Results remain controlled: the evaluation uses motion-capture support, omits a physical net, and does not demonstrate strategic shot placement or unconstrained perception. [Technical summary](https://www.alphaxiv.org/abs/2609.31840)