Real-time Modelling of Complex Dynamical Systems

Proceedings of the TAROS (Towards Autonomous Robotic Systems) intl. conference, September 12-14, 2005, London

Christfried Webers & Uwe R. Zimmer

Dynamical systems modelling for physical, real-world systems is a recurring task in many robotics applications. It is often preferable or even required that the creation/adaptation of such a model is performed on-line with the actual stream of sensor samples and also under realistic realtime constraints. This article suggests a new method for real-time modelling of complex dynamical systems, overcoming the "curse" of dimensionality which applies to any complete modelling system known up to now and which prevented any real-time implementation in even moderately complex environments.
The key concept of the proposed method (RTDSM) is to define a trade-off between the dimensionality of the underlying delay-vector space and local delay-vector statistics (called "histories "). Combined with locally limited adaptations in each step and the exclusive employment of each sensor sample once and in sampling order, the method recommends itself for on-line and real-time applications, while keeping the modelling power of off-line analysis methods. Restrictions apply, and so the understanding of the trade-off introduced by this method is crucial for its successful application. Simulations and a comparison to an established related method are given.

Paper (pdf) | other publications