Seminar: N. Mansard, "Task function for dynamic animation and recognition", Sept 20 at 11:00
Speaker: Nicolas Mansard, LAAS-CNRS, Toulouse, France
Title: Task function for dynamic animation and recognition
Time and location: Tuesday, September 20, 2011, at 11:00, A3
The task function is an elegant approach for describing a motion: instead of working in the configuration space, a dedicated "task" space is specifically built to represent the objective, where the control law is easy to define. Then, converting the control law from the task space to the whole-body control space is simply a matter of numerical computation. After having recalled the classical basics (classical redundancy, pseudo inverse, no inequality constraints), I will discuss the composition of multiple tasks using equalities and inequalities to control the dynamics of the system. This enables to obtain nice dynamical movements, comprehending both the balance and the full velocity of the system. However, several problems still remain, e.g., computation cost, numerical conditioning, and balance stability. A second use of the task function approach can be as a descriptor of the motion: it can then be used to recognize a motion, using a descriptor language that enables straightforward imitation.
The second part of the talk will present a first solution for task-based motion recognition, and its use to disambiguate similar-looking motions without using the context.