Reproduce “Beyond Imitation: Generative and Variational Choreography via Machine Learning” on OpenResearch

About this paper

Mariel Pettee, Chase Shimmin, Douglas Duhaime, Ilya Vidrin

Researchers from Yale University and Coventry University developed machine learning frameworks that generate novel dance sequences and tunable variations of existing movements using high-dimensional 3D motion capture data. Their LSTM-MDN and autoencoder models provide choreographers with tools for creative exploration and self-reflection on embodied knowledge, producing smooth and authentic-looking animated outputs.

Key insights

• Machine learning tools can effectively augment choreographic practice by offering generative movement 'playgrounds' and fostering introspection into embodied knowledge, rather than replacing human creativity.
• Ensuring continuity and smoothness in the latent spaces, achieved through autoencoders and preprocessing steps like removing global orientation, is crucial for synthesizing meaningful and physically coherent choreographic sequences.
• Controlling the amount of noise added to a VAE's latent representation allows for precise control over the degree of variation in generated choreography, ranging from subtle stylistic changes to inventive deviations.

Summary by alphaXiv.