ChoreoPlan: Hybrid Phrase Planning and Execution-Grounded
Selection for Music-to-Humanoid Dance

Abstract

Music-to-humanoid dance is commonly implemented as a two-stage pipeline in which a music-conditioned generator produces a reference motion and a fixed whole-body controller (WBC) executes it. In this setting, reference-space quality is only a proxy for the quality of the executed motion. We focus on two upstream decisions that strongly affect this proxy gap: how temporal planning units are defined and how generated candidates are selected before execution. ChoreoPlan introduces beat-snapped, variable-length planning segments with hybrid discrete and continuous motion attributes, together with an Embodied Selector trained on offline rollouts of the fixed controller. The planner provides beat-aligned choreography guidance in humanoid token space, while the selector reranks candidates using predicted execution quality and semantic compatibility. Across AIST++- and FineDance-derived humanoid tracks in IsaacGym and MuJoCo, ChoreoPlan improves rollout success, tracking accuracy, executed beat alignment, and semantic retention over retrained baselines. Qualitative Unitree G1 demonstrations further illustrate coherent and executable dance motions.