Distributed neural control for complex autonomous robots

Chapter 1: Overview


This thesis focuses on the creation of behaviours in complex robots using artificial neural networks (ANNs) as information processing elements in a distributed neurocontrol architecture. The term ”complex robot” is understood as a physical agent composed of a large number of sensors and actuators. The generation of neural controllers for these robots is a complex task due to network training, because in most cases training sets for the task at hand are not available, and even when examples are available it is still unclear as to precisely how to allocate error to the different components involved in the control of the robot for the given task.

One possible approach is the use of evolutionary robotics methodology for network training. Evolutionary robotics (ER) uses evolutive algorithms for network weights updating [Nolfi and Floreano, 2000], avoiding training examples and blame assignment. The use of ER however, presents its own set of problems; it requires real interaction between the robot and its environment, and it does not scale well as the complexity of the robot and/or its behaviour grows up. This thesis will deal with these problems by developing a general distributed control architecture for robots. This architecture is based on neural networks as basic processing elements, and will be independent of the task, the environment, and the robot. The architecture is therefore proposed as a solution to the problem of generating behaviours in complex robots within the evolutionary robotics paradigm.

Figure 1: A simple robot (left, the Boebot) and a complex one (right, the
Aibo robot). A simple behavior to move around those robots, like for example,
a wall following behavior, implies, in the case of the Boebot robot, the evolution
of a controller for 4 devices (sensors and actuators). The same behavior for the
Aibo robot implies the evolution of a controller able to handle between 25 and
30 devices (depending on implementation).

The main problem to solve

When the robot to be controlled is complex in terms of number of sensors and actuators, the search space for the evolutionary algorithm becomes immense, and the algorithm is unable to find a solution within a reasonable amount of time. Moreover, when the controller to be evolved is complex in terms of the task, the complexity of the search space may become so high that it could become impossible to find even an initial partial solution with which to guide the evolutionary process toward the final solution.

The thesis work

In this thesis, a solution is proposed to these types of problems. A distributed architecture for the control of complex robots using ANNs has been created. It makes use of evolutionary techniques to find the most suitable network weights. The distributability of the architecture makes it possible to partially overcome the curse of the dimensionality problem usually encountered in evolutionary processes. The architecture is of a general purpose in the sense that it is independent of the robot and the task to be solved. The architecture is of the reactive type. Additionally, the architecture allows for the coupling with a deliberative process.

different applications of the DAIR architecture
Figure 2: The DAIR architecture, described along the thesis, applies to any
type of robot, independently of its number and type of devices.

Related published papers

O. Vilarroya and R. Téllez, La madurez de los Aibo (in Spanish), Palabra de robot, Publicacions de la Universitat de València, 2006

R. Téllez, Robots i evolució (in Catalan), in ACIA Newsletter, 2004. Winner of the honorable mention award at the 1st ACCC competition for the spreading of scientific research, 2003.

Web page by R. Téllez using rubric css by Hadley Wickham
Don't undertake a project unless it is manifestly important and nearly impossible (Edwin Land)