This page gives an overview of my projects related to the visual control of autonomous robot systems.

Robot systems that are to operate in a unknown or changing environment need a means to perceive the environment and how its actions change it. This interaction can be modelled by a Perception-Action Cycle, or for short: PAC. Our research is focused on ways to implement these PACs to develop "intelligent" robot systems. Our methods combine approaches from engineering (feedback control), mathematics (control theory and optimisation) and computer science/biology (neural networks and evolutionary methods).

In this project an Image-based Visual Servoing robot controller was designed, implemented and validated. In order to be able to manipulate an object the controller's task is to move the robot end-effector into a desired pose relative to the object. The location of the object in robot space is unknown. Only image data from a CCD camera attached to the robot's end-effector (eye in hand) is used to calculate robot movements as controller output. Image features corresponding to the desired pose were determined by moving the robot there and acquiring an image (teaching by showing). The object carries a label with 4 markings. The difference between their actual and desired positions in the image (image error) is the only controller input. The implemented algorithm is of the type image-based static look and move. Because of their robustness towards model errors, applications for image-based robot controllers include autonomous robot systems. An example is the University of Bremen's FRIEND system to support disabled people where a version of this controller has been ported to.

A serious problem of previously implemented controllers is the handling of errors in the system model, particularly external camera parameters. Especially with an eye in hand camera, resulting large inaccurate robot movements can produce undesired effects. Apart from the danger of hitting objects or people inside the robot workspace object markings are often no longer visible in the image and the controller does not converge. Using a Trust Region Method to determine its output, the newly developed visual servoing controller prevents these problems. This is done by measuring model errors and automatically adapting a maximum step length for the controller. By taking as long steps as possible while maintaining convergence the new controller guarantees a successful visual servoing process. At the same time the number of steps required to move the robot into the desired pose is kept very small.

Yohannes Kassahun has introduced the method "EANT", "Evolutionary Learning of Neural Topologies" in his PhD work in our research group during the time 2003-2006. Since then it has been further developed and applied to learning neural controllers for visual servoing.

More details about the developed method, EANT, can be found on this page on evolutionary learning of neural networks with EANT.

The European Framework 6 Project COSPAL (July 2004 to June 2007) is concerned with the development of a system design for systems which combine perception and action capabilities to solve complex planning and manipulation tasks. Methods using supervised, unsupervised and reinforcement learning are employed. An important focus is also on incremental learning such that the system can adapt to new situations and tasks during its lifetime. The learning methods are based, on Artificial Neural Networks (e.g. DCS Networks with online learning capabilities), Associative Networks and others.

For the development and testing of the newly developed methods we are implementing a demonstrator system that solves a shape-sorting puzzle like the ones used as toys. The system has several layers of abstraction, with the highest being a symbolic processing unit. At each layer a Perception-Action Cycle (PAC) can be identified, building a hierarchy of PACs. While traditional methods (or a 3-year old child) could be used to solve the shape-sorting task it is our goal to use only methods that learn how these PACs are implemented.

The task of our group, the Cognitive Systems Group of the Christian-Albrechts-University of Kiel, is to implement learning methods for those parts of the system closest to the robot system hardware. The perception/action modules and controls learned by our group are those for detecting objects and extracting features from images, classifying these features and moving the robot (long range movement, alignment, obstacle avoidance). An emphasis is also placed on learning to imitate human-like movement while solving the task.

More information can be found on the COSPAL project home page.

This page and all files in these subdirectories are Copyright © 2004-2010 Nils T Siebel, Berlin, Germany.