Section: Contracts and Grants with Industry
Keywords : Bayesian Model, Behavior Modeling and Learning.
ROBEA Bayesian Models for Motion Generation
Participant : Stéphane Donikian.
The ROBEA (CNRS Interdisciplinary Research Program) Project entitled "Bayesian Models for Motion Generation" is a partnership with the Cybermove and Evasion Research Projects of the Gravir Lab in Grenoble. The aim of this program is to study how bayesian models can be used to teach an autonomous agent its behaviors, instead of specifying all the probability distributions by hand. It requires to be able to measure at each instant sensory and motor variables of the controled character. The first year has been mainly devoted to the integration of the bayesian programming and interactive natural sceneries modules developed by our partners inside OpenMASK. In the second year, we have developed the urban application that will be used to study the learning by example of a pedestrian navigating in a virtual city. In this third year, we have studied how bayesian programming can be used to learn a behavior by example and we have made some navigation experiments.
Parameters used to learn the navigation task of a character consist in its speed and the distance and orientation of the next cell boundary in the path to be followed. Each parameter is described by a set of possible discrete values. The speed vector module is linearly discretised in five values [0..4] in the interval from 0 to 2 m. s^{-1} . The speed vector orientation is discretized in three value: -1 if the angle is greater than /24 , 1 if the angle is lower than - /24 , and 0 in the other case. The orientation of the target point is discretized in nine values in the interval [-4..4] based on the cubic function ( x^{3} in the interval [-1..1] , while its distance is discretized in five values [0..4] in the interval from 0 to 40, based on the quadratic function x^{2} in the interval [0..1] .