From the perspective of Evolutionary Robotics, evolutionary algorithms have been suggested as a developmental tool for achieving adaptive robots. At present, there is much discussion as to which type of robot control system to evolve and whether to use simulations of real robots. There are examples of evolved controllers based on explicit programs in high-level language, implemented as classifier systems, fuzzy systems, and a large variety of neural networks. Further, there is some initial work towards the combination of traditional learning algorithms (at the individual level) with the evolutionary algorithm (on the population level). Some researchers have made use of experiments with whole evolution processes running on-line, while others have used different kinds of simulators before transferring the evolved controllers to real robots interacting with the real world. The main problem with the on-line approach is the enormous time consumption of running evolutionary processes on-line for many generations, while the main problem when using simulators is to reduce the gap between performances observed in the simulated environment and performance obtained in the objective environment. From an engineering perspective, we would like to observe the same behaviors in both environments. From a more biologically plausible point of view, the problem is how to produce, along phylogenetic evolution, artificial genotypes that adapt, in the course of onto-genetic evolution, to different changing environments. In this perspective, Evolutionary Robotics uses the contributions from a wide range of disciplines and it represents an empirical field where testing hypotheses about the mechanisms of natural evolution at different levels of analysis is made possible. Evolutionary Robotics can, for example, find its place in a neuroethological context, where the robotics approach is used to verify the capabilities of known, natural control systems, and as a study of artificial life, Evolutionary Robotics leads to interesting new life forms. Evolutionary Robotics can be used to study the interaction and limitations that different robot parts constitutes. The evolution of a controller is limited by the body plan of the agent that it controls. In the case of robots, the available sensors, motors, wheels, etc. put a limit on the possible controller to evolve. However, some research in Evolutionary Robotics suggest, that both controller and robot body plan can and should co-evolve (as is true for natural organisms). From an engineering point of view, this Artificial Life technique gives a clear advantage over traditional robotics techniques, since it allows an automatic construction toward a goal-fulfilling behavior and robot structure. The use of Evolvable Hardware is somewhat on the same line of research. Here, an evolutionary process changes reconfigurable electronic circuits (i.e. control systems of a robot) to adapt the characteristics of the specific robot and environment. The evolution process reconfigures the hardware directly, so the constraints imposed by the hardware are satisfied automatically. The resulting systems often seem extraordinary to engineers, but have proven to be of very high power and efficiency. An interesting development of this is into evolvable biochemical systems such as the micro-integrated fluid system. Such systems are more biological plausible and integrated with robotics they might provide new knowledge about controller/body structure interaction, especially when combined with biological sensors.
Evolutionary Robotics is an appealing idea from the Artificial Life community, but it needs much experimental investigation, since there are still open questions, for example regarding its scalability. Will monolithic systems whose evolution is based on a single, global fitness measure scale up to more complex behaviors? Is modularization with definition of fitness criteria for each single behavior a necessity? How is a modular system to evolve without demanding too much a priori task knowledge? This special issue of the Artificial Life journal should provide both quantitative data, empirical techniques, and standards for researchers in Robotics and Artificial Life.
The following list suggests some topics of interest, though it should be viewed only as a guideline.
http://mitpress.mit.edu/journal-home.tcl?issn=10645462
the only exception being, that submission should be made via anonymous ftp to
ftp.alife.org:/journal/robotics/incoming
(or four hard-copies sent to one of the guest editors). Send an e-mail notification to one of the guest editors after uploading! Look at the ftp instructions.
Henrik Hautop Lund
Minoru Asada