Abstract:
This thesis addresses the fieid of visual servoing and deals with solution nnethods for the non-adaptive estimation of the Jacobian in learning approaches for visual servoing. Sensors are taken as abstract sources of Information that are used to identify robot movements by linearizing the coherence between sensor data deviations and robot position deviations in a Taylor expansion with a Jacobian. We compare different methods to determine the Jacobian from learning data such as the inversion of the Feature Jacobian or the direct solution of an over-determined system, regarding the properties of practicai relevance for visual servoing appiications like the flexibility of the calculation and the quality and the stability of the result. In the conclusion, we are able to suggest a procedure to extract the Jacobian out of learning data. The analyzes are illustrated by multiple examples from real worid experiments.
Another part of this work deals with an effect that occurs when collecting learning data; a non-linearity of hysteresis-type in the samples of robot positions and the sensor data. This effect can cause problems when identifying the system with some of the previousiy mentioned methods. We explain the origin of this effect (overlay of sampling delays and mechanical robot effects) and discuss methods for the determination and compensation of the effect in the learning data.
