Motion Representation Learning

Motion representation learning studies how to encode movement from sources such as body skeletons, wearable sensors, point trajectories, and human–object interactions so models can recognize, predict, generate, or control motion. Approaches range from discrete motion tokens and continuous latent spaces to geometric, graph-based, and frequency-aware representations, trained with masked prediction, contrastive learning, or temporal prediction. Language and video supervision increasingly connect these encodings to semantic descriptions and robot actions, while diffusion and flow-matching models use them for tasks such as text-to-motion generation and motion completion. A central challenge is balancing compact, efficient representations with fine-grained dynamics and meaningful semantics, while ensuring transfer across datasets, body shapes, sensor setups, and real-world conditions.