|
|
|
Abstract:
|
Designing controllers for robotic applications can be very difficult due to the nonlinear nature of the involved dynamical models and certain lack of knowledge about their operating conditions. This is especially true for autonomous mobile robots that have to cope with unexpected real-world events. In this thesis I present a systematic way of including learning control on top of the traditional model-based control to improve the overall performance of the system. The inspiration for this research arises from one type of biology’s animal classification denoted altricial-precocial spectrum. In nature, some animals are born very well developed (precocial), while others are under-developed and lack any complex skills at birth (altricial). This classification finds an analogy in control theory where one can find model-based controllers on one side and learning algorithms on another. The objective is to combine controllers at both ends of the altricial-precocial spectrum for robotic motion control, thus taking advantage of the accurate understanding of model-based controllers and the ability of learning for coping with unmodeled properties and unexpected events. In this thesis, the application of a supervised machine learning method called Local Linear Regression (LLR) as learning control is investigated. This learning method “remembers” the experiences of the robot explicitly and they are stored in a manner which permits fast recall of the closest previous experience to any new situation. The application of the Altricial-Precocial control framework is demonstrated on a DC motor setup as an online gravity compensation method to improve steady-state performance. It is also applied on a legged robot to automatically adjust the gains of the original controllers in order to improve their performance due to friction and other dynamics while walking around.
|
