Unified Human Motion Pretraining via Generative Chrono-Structural Decomposition and Alignment

3D human motion-language alignment remains challenging due to the complex entanglement of spatiotemporal semantics. Previous methods typically rely on holistic alignment, which overlooks fine-grained structural primitives, i.e., body parts and sub-events, inevitably leading to imprecise correspondences and limited generalization. To address this, we propose MoLA, a unified motion-language pretraining framework that explicitly decomposes and aligns motion/language at chronological and structural levels. Leveraging large language models (LLMs) as prior knowledge experts, we reformulate the coupled textual descriptions into structured part- and event-level motion concepts, enabling precise supervision for fine-grained alignment. Based on this, we further develop novel chrono-structural alignment mechanisms to facilitate correspondence learning. At the spatial level, we introduce a part-mixture learning strategy that decouples body-part motions to facilitate part-decorrelated correspondence modeling. At the temporal level, we propose Gumbel-based adaptive consistency learning coupled with a chronological order objective to address this weakly supervised event alignment task. This enables the model to dynamically localize relevant motion clips and learn temporally coherent correspondences without requiring explicit boundary annotations. Furthermore, a directional consistency regularization is involved that models semantic relationships between global motions and their local patterns. Extensive experiments on motion-text retrieval, recognition, captioning, and generation demonstrate consistent improvements over state-of-the-art methods, verifying our motivation and effectiveness for unified motion pretraining.