Knowledge-Driven World Modeling in Robotics
|has title::Knowledge-Driven World Modeling in Robotics|
|Master:||project within::Computational Intelligence and Selforganisation|
|Student name:||student name::Sjoerd van den Dries|
|Supervisor:||Maria Gini, Mark Hoogendoorn|
|Company:||has company::TU Eindhoven|
In the RoboEarth project, a multi-disciplinary partnership of robotics researchers aims to create a World Wide Web for robots: an online database through which robots can share learned knowledge and experiences. Such a repository would enable robots to benefit from the experience of other robots, which could greatly speed up the learning and adaptation process that allows robotic systems to perform complex tasks. One of the key challenges in this enterprise is to find a way of representing knowledge that 1) is both expressive and generic; 2) can be learned from experience; and 3) can be shared across different platforms and different types of hardware. Besides, a system should be defined that is capable of dealing with this type of knowledge in a controllable and dynamic way.
My research aims to provide such an interface and implementation for the module that performs data association and tracking (world modeling) on AMIGO, the robot that is used by Eindhoven in the RoboCup@Home competition. This module is implemented using a Bayesian Multiple Hypothesis Filter, a structure that keeps track of the most probable hypotheses about the state of the world using a specified probability model. By adding object and context specific knowledge to this probability model, the performance of the filter can be improved, resulting in a more accurate model of the world. In the end, the robot should be able to learn the probability model himself, and share this type of knowledge through RoboEarth to allow other robots to benefit from it as well.