Diffusion Models for Motion Generation
Overview
This field studies how to generate, edit, and forecast movement, from human gestures and dance to robot actions, crowds, and interactions with objects or scenes. Common approaches use conditional denoising in motion or latent spaces, with transformers, variational autoencoders, autoregressive or streaming generation, and flow matching; conditions can include language, music, poses, paths, images, and physical constraints. These methods can turn flexible instructions into motion for animation, simulation, and embodied control, and are commonly evaluated on benchmarks such as HumanML3D, KIT-ML, and AIST++. A central challenge is balancing text or task control, natural and physically executable motion, long-horizon coherence, and fast inference, while ensuring benchmarks reflect real-world performance.
Latest papers
98 papers20 papers in the last four weeks, up 150% on the four weeks before. 0.1% of all new papers.