Unification of Task Identification and Control for a Humanoid Robot

We present a method using simple mechanisms to perform a real time imitation on a humanoid robot. Instead of focusing on the fidelity of the motion, we will rather focus on what controllers have to be activated to perform a desired motion. The originality of our approach is to work in the task spaces of the robot in order to solve simultaneously the identification and the execution of a motion. With this approach, we bypass the classical phase of the transformation from the imitated model to the imitant model. The tasks identification method consists in proposing several model of the motion, and then computing the parameters of this model with a least square optimization. The motivation behind the use of a simple technique is to be as much efficient as possible regarding computation time, as this work is a preliminary step for the resolution of the reactive control problem for a humanoid robot where more complex mechanism have to be used. We formulate it as the problem of deciding very quickly which sequence of controllers, has to be activated (this choice will be based on motion imitation) to realize a motion. Although in this work we consider the problem where a robot imitate a robot, we investigate the possibility of extending our method to sequential motion imitation and a case where a robot imitate a human.