Learning Motion Patterns from Multiple Observations along the Actions Phases of Manipulative Tasks

Abstract

In this work we present a probabilistic approach to find motion patterns in manipulative tasks by looking for similarities among the relevant features along of the actions phases of a trajectories dataset. From multiples observations of human movements we can align all signals temporally to perform a learning process based on selection of relevant features by analyzing their probability distribution and finding correspondent features with high probability in each phase of the trajectories of a dataset. Using the spatio-temporal information of the learned features we can generate a generalized trajectory of the dataset using a polynomial regression to fit the features data by successive approximations. The smoothed trajectory can be used as a prototype/template for matching (1:1) or for classification (1:N) using Bayesian techniques to know if a new observation is similar to a specific task or to recognize a task. The intention here is to have an approach that is able to learn and generalize a specific movement by their patterns to be applied in the future for different contexts. We are not going through the imitation learning part, but we are focusing on the ability of learning to reach some intelligence to approximate a movement generalization, tasks that humans do in a natural and easy way.