Année de soutenance: 2011

Manuscrit accessible ici

Résumé

Efficient methods to perform motion recognition have been developed usingstatistical tools. Those methods rely on primitives learning in a suitablespace, for example the latent space of the joint angle and/or adequate taskspaces. The learned primitives are often sequential : a motion is segmentedaccording to the time axis. When working with a humanoid robot, a motion can bedecomposed into simultaneous sub-tasks. For example in a waiter scenario, therobot has to keep some plates horizontal with one of his arms, while placing aplate on the table with its free hand. Recognition can thus not be limited toone task per consecutive segment of time. The method presented in this worktakes advantage of the knowledge of what tasks the robot is able to do and howthe motion is generated from this set of known controllers to perform a reverseengineering of an observed motion. This analysis is intended to recognizesimultaneous tasks that have been used to generate a motion. The method relieson the task-function formalism and the projection operation into the null spaceof a task to decouple the controllers. The approach is successfully applied ona real robot to disambiguate motion in different scenarios where two motionslook similar but have different purposes