Guided latent space regression for human motion generation

In the present work, we describe a mathematical model to generate human-like motion trajectories in space. We use linear regression in a latent space to find the model parameters from a set of demonstration examples.The learning procedure requires a relevant set of similar examples. The apprehended models encode both the typical shapes of motion and their variability towards specific boundary conditions (BC). We will show the added value of encoding both properties in a unique model and we apply this ability to common problems of error compensation and target tracking.The models allow us to describe human motion using expansion-function series (EFS), thus avoiding typical stability issues that arise in the use of differential equation models. To cope with variable scenarios, we show two specific algorithms that morph and adapt the evolution trajectory. In analogy to splines, the EFS preserve an analytical structure on which we develop the optimisation steps. In such a way, we managed to combine multiple single segments into complex motions that preserve continuity and may simultaneously optimise other criteria.In the present work, after having analysed similar tools, we present the basic model and its features. Then we develop a robust tool to gather the model from examples, and to achieve real-time trajectory adaptation. The achieved results will be analysed through an experimental analysis on data collected in a ball catching experiment.

► A latent space regression to estimate human motion.
► Model based on discrete sine transformations and expansion-function series.
► Versatile models that can be assembled to recreate any complex motion.
► An automatic training tool that learns both motion and its space properties.
► Tracking tools to match boundary constraints and to adapt motion shapes in real-time.