Learning Methods for Autonomous Mobile Robots

Alexander Gloye

The dissertation describes learning algorithms for autonomous mobile robots. 
Different learning methods are applied to various levels of complexity and 
abstraction of the robot in order to enhance its behaviour and to accelerate the 
development of the whole system. The presented techniques vary from the on-board 
feedback control of the motors to an automatically learned simulation system.  

This thesis introduces approaches for compensating the dead-time of a system by 
observing and learning its behavour. Based on a learned prediction model, 
different methods are developed to
* control a robot with a defect motor,
* optimize the driving pattern of the robot on different levels of abstraction,
* learn the driving dynamics without a physical model, and
* eliminate control failures and errors.

The specified learning methods can simply and safely be applied to other robots 
and plants. The presented techniques have the common property that they simplify 
and accelerate the development of robots. Time-consuming analytical processes 
for the design of physical models are thus omitted, because the algorithms learn 
and optimize the parameters and models of the robots automatically.