@techreport { yin18, abstract = {In this paper, we learn dynamics from highdimensional demonstrations to facilitate model-based prediction and robot control. The proposed approach leverages the progress in variational-bayes and sequence modeling, extracting a low-dimensional latent space so the dynamical relations of interest can be compactly represented and learned. Different from existing works, our model captures latent dynamics in a more general form and features efficient inference for pattern filtering, prediction and synthesis. The extracted feature mapping and latent dynamics can be naturally integrated in robot learning, yielding task imitation from raw data and prediction-based reproduction. The performance of latent dynamics learning and model-based imitation is shown in three tasks: 1) reconstructing and predicting images of bouncing balls movement with an accuracy competitive to the state-of-the art; 2) synthesizing diverse handwriting image sequences; 3) learning to strike a ball under partial visual input, with results significantly outperforming baselines.}, institution = {Intelligent Agents and Synthetic Characters Group}, keywords = {Miscellaneous;Neural Computation;}, month = {May}, number = {GAIPS-TR-001-17}, title = {Learning Variational Latent Dynamics: Towards Model-based Imitation and Control}, year = {2018}, author = {Hang Yin and Francisco S. Melo and Aude Billard and Ana Paiva} }