Francesco Mondada
There is a large number of possible applications in the field of mobile robotics: Mail delivery robots, domestic or industrial vacuum cleaners, surveillance robots, demining robots and many others could be very interesting products. Despite this potential market and the actual technology, only few simple systems are commercially available. This proves that there are several important and problematic issues in this field, mainly at the intelligence level. As a reaction to the failure of the classical artificial intelligence applied to the field of mobile robotics, several new approaches have been proposed. Artificial neural networks are one of these, and genetic algorithms, supported by the Artificial Life trend, are also getting more and more consideration. These two techniques have already been applied to mobile robotics, but mainly in simulation, and without a final test on a real mobile robot. The use of physical robots for this research seems to be still problematic due to the lack of efficient tools. Several neural structures for the control of mobile robots have been analysed in this work. All experiences have been carried out on physical robots. To reach this goal, an important effort has been made in order to design new efficient robotic tools. Together with Edo Franzi, André Guignard and Yves Cheneval, we have developed and built hardware and software tools that make an efficient research work possible. Along with several analysis software tools, the mobile robot Khepera has been a result of this development. Using this equipment, six experiences have been carried out, covering a large spectrum of the possible ways neural networks can be used for the control of mobile robots. These experiments have nevertheless been restricted to simple behaviours and small neural networks. The first two experiments show, with a very simple and manually adjusted behaviour, the important role of the interaction of the robot with its environment. The first experiment is based on a collective behaviour, the second on a collaborative one. The adaptation of the robot to the environment is introduced in the third experiment, in which a learning technique is applied. The result is a robot able to learn how to use visual stimuli to avoid particular obstacles. Despite its interesting results, this approach has turned out to be very limited, due to the rigid structure needed. The last three experiments demonstrate the possibilities of the use of genetic algorithms, which proved to be a very flexible adaptation mechanism. The first of these three experiments tests the feasibility of this approach. The second one takes advantage of the characteristics of genetic algorithms to achieve more complex behaviours. Finally, genetic algorithms and learning techniques are associated in the last experiment, showing a high adaptive structure. An important effort has been made to show both advantages and disadvantages of each technique, in order to provide the necessary elements for the continuation of this research activity.
EPFL1997
