Ultra fast cnn based Hardware Computing Platform Concepts for adas visual Sensors and Evolutionary Mobile Robots

 
98 
presented in this chapter an irregular and even a disjointed walker robot can successfully 
move with the highest performance.
9.3
 
Introduction
 
Nowadays, some of the main goals of robotics science, mechatronics and artificial 
intelligence lie in designing mechanisms close to or mimicking as good as possible some 
natural structures or animal behavioral models. According to this theory, the nature selects 
the powerful and stable genes for breed, and weak genes fall/disappear in the nature 
[141]. The good genes that can adapt the animal structure to the environment have higher 
chances for breed and evolution. The animal locomotion is trained and adapted according 
to the animal’s body structure. One key issue in the training process is based on the energy
saving. This justifies the striking interest devoted to the modeling and simulation of animal 
walking motion with the aim of optimizing the energy consumption [142-144]. It is well-
known that the walking motion of animals is of a stereotype. In a large variety of animals a 
central neural controller does organize/coordinate the motion. A central neural controller 
(e.g. the 
Central Pattern Generator
(CPG)) is a main unit for controlling limbs for walking 
[145]. The CPG unit does contain all the mechanisms needed to generate the rhythmic 
pattern of movement. This unit is suitable for designing walker, swimmer, or flyer robots 
which exhibit motion close to natural locomotion mechanisms. Due to recent advances in 
electronics and the ability of cellular neural networks to solve partial differential equations 
in real time, it is possible to simulate a Reaction-Diffusion model by a specific CNN 
architecture, the so-called 
Reaction-Diffusion Cellular Neural Network
(RD-CNN).
A striking interest has been devoted to the robot control based on the RD-CNN technique 
[141, 145, 146]. In this technique, the mathematical model describing the robot behavior 
must be well-defined. This is a serious limitation as modeling the complex behavior of 
robots is challenging. In this chapter we introduce a robots control method which is not 
based on the mathematical modeling of the robots behavior. It is rather a general and 
effective method combining CNN with GA. This method can support and drive many types 
of structured and unstructured walker robots.
The method/approach is based on both the 
natural modeling and the use of computational units close to biological models. A 
combination of both CNN (i.e. for computation) and GA (i.e. for optimizing the nature) is a 

 
99 
good tool for modeling and controlling robots dynamics. The central parts of this scheme 
are made-up of a CNN processor and an evolutionary training unit. A 
Cellular Neural 
Network
(CNN) is a parallel computing paradigm similar to the artificial neural networks 
computation platform, with the difference that in CNN the communication is allowed 
between neighboring units. This feature of the CNN processor makes it a good computation 
platform to analyze the dynamics of biological neurons. This research shows the possibility 
of directly driving a walker robot by an evolutionary training of a CNN processor. This 
method is further efficient to model widespread natural locomotion mechanisms of 
animals (e.g. worms, insects, quadrupeds, biped, etc) [104]. This locomotion is modeled in 
the 3D space describing the real environment and in very difficult situations (i.e. rough, 
bumpy, and/or scaly surfaces) as well. The challenging focus is finding the best signal for 
driving walker robot joints with minimum energy consumption and the best locomotion 
performance. This can be achieved by finding suitable CNN templates to generate an 
efficient wave for driving the walker robot joints. This chapter is organized as follows. 
Section 9.4 discusses the use of genetic algorithms for optimizing the CNN templates.
Section 9.5 presents the training algorithm and some simulation results as well. Section 9.6 
formulates some concluding remarks. Further, the quintessence of the results obtained is 
summarized, and some open research questions are outlined.

Download 3,22 Mb.