MotionVerse: Toward Unified Human Motion and Interaction Modeling

Existing motion-language frameworks primarily focus on isolated tasks such as single-person motion generation or understanding, while more complex scenarios, including human interaction, reaction, and motion editing, are typically addressed using task-specific architectures. Despite sharing common motion semantics and temporal structures, these heterogeneous behaviors are rarely unified within a single modeling paradigm, limiting transferable motion reasoning and general-purpose motion intelligence.

To address this limitation, we present MotionVerse, a unified motion-language framework for modeling both individual and interactive human behaviors. MotionVerse adopts a scalable multi-stream autoregressive formulation for residual motion tokens, enabling efficient long-sequence modeling while preserving fine-grained motion dynamics. To alleviate modality interference during joint optimization, we further introduce a modality-specialized Motion-Fusion mechanism with dedicated adaptation pathways for motion and language modeling.

Extensive experiments across diverse motion-language tasks demonstrate that MotionVerse achieves competitive performance in motion understanding, generation, interaction, reaction, and editing, validating its effectiveness as a scalable and general-purpose framework for unified motion modeling.