Continuous humanoid badminton rallies and full-court coverage

Develop a humanoid-robot badminton control framework that enables continuous human-robot rallies and scales locomotion to support full-court coverage, extending beyond single-shot interception within a localized area.

Background

The paper introduces a progressive Imitation-to-Interaction framework that learns human-like badminton strikes on a humanoid robot, demonstrating zero-shot sim-to-real transfer for skills such as forehand and backhand lifts and drop shots. The method builds a motor prior from human demonstrations, distills it into a compact representation, stabilizes it with adversarial motion priors, and finally refines it through physics-based interaction.

Despite achieving accurate single-shot interception in localized regions, the authors explicitly note that enabling sustained rallies and traversing the entire court remain open challenges. Addressing this problem would require integrating robust whole-court locomotion with agile, timing-critical striking while maintaining stylistic naturalness and stability under real-world constraints.