В библиотеку
Evolutionary Robotics: The Next-Generation-Platform for On-line and On-board Artificial Evolution
Kernbach S., Meister E., Scholz O., Humza R., Liedke J., Ricotti L., Jemai J., Havlik J., Liu W.
Abstract
In this paper we present the development of a new self-reconfigurable robotic platform for performing on-line and on-board evolutionary experiments. The designed platform can work as an autonomous swarm robot and can undergo collective morphogenesis to actuate in different morphogenetic structures. The platform includes a dedicated power management, rich sensor mechanisms for on-board fitness measurement as well as very powerful distributed computational system to run learning and evolutionary algorithms. The whole development is performed within several large European projects and is open-hardware and open-software.
Introduction
Currently, the evolutionary paradigm represents a new research trend for adaptive, reconfigurable and collective robotic systems [1]. These systems are primarily characterized by their ability to reconfigure autonomously, where the robots can work either individually or aggregate autonomously into collective robotic organisms [2]. The application of an evolutionary approach for these systems enables adaptation to new, even previously unknown, environments by evolving an individual or cooperative behavior (e.g. [3], [4]). In particular, a collective actuation of robotic organisms, being a complex function of many interacting individuals, can be obtained in the evolutionary way [5], [6]. Here, we think about several relevant long-term applications, where a continuous communication with a human operator is impossible, like deep space or ocean explorations, robotic Mars missions or other human-less environments. However, working on such systems, we face a few serious fundamental issues needed to be considered.
Firstly, such robotic systems operate either in fully autonomous mode or with intermittent human corrections of strategic goals. Consequently, the learning and evolutionary processes should be executed only by those means which are available on board. This is related to an acquisition of sensor data, fitness estimations, computational and communication resources. Secondly, the robotic evolution should succeed within operative time-scale of a concrete environment. Usually, it means a time interval between several minutes and several hours. We consider these properties necessary for on-board and on-line evolving. Finally, several essential issues appear due to reconfigurability of the system. Since a collective functionality depends on the morphology of robotic organism, the evolution process of a collective functionality implies a collective morphogenetic process. This, in turn, poses many open questions regarding the relationship between functions and structures [7], between individual and collective fitness, between learnt and evolved behavior.
Hence, for the development of real evolutionary adaptive systems, we need a robotic platform which can provide several specific capabilities. First of all, it should be selfsufficient in order to support self-reconfigurable capabilities with individual/collective actuation and with autonomous docking mechanisms. Moreover, it should possess enough computational power and sensing capabilities. Finally, it should be reproducible by interested research groups with a reasonable amount of effort. The last point is closely related to the following issue.
Developing approaches and mechanisms for such evolutionary reconfigurable systems, we faced the problem, known as a “gap of reality”. Since on the early stage of projects no real hardware was readily available, simulations were used allowing a preliminary software development. However, these simulations do not provide the complexity of a real environment, resulting in simplified sensor data, which do not reflect many real non-linear phenomena (like reflection of IR-light around proximity sensors). Moreover, deriving a fitness functions in simulative environments, we often encounter a “linearity” of fitness. The “linear fitness” does not consider some specific “ecological fitness niches” existing in a real fitness landscape. As a result, such evolutionary solutions, which are obtained in reality, are not accessible in simulation [8]. As a reaction to the “gap of reality”, we developed the “hardware simulator”, a composition between real hardware and computer simulation, which is easily available for the software community and provide a realityclose evolving of robot controllers.
In this paper we represent the development of such a selfreconfigurable platform for evolutionary community. This development is conducted by a consortium of twenty research organizations within the European projects “SYMBRION” [9] and “REPLICATOR” [10]. The development as well as the platform itself are intended to be open-hardware and open-source. Moreover, on the early developmental stage, the consortium is collecting ideas and requirements from the areas of evolutionary computation, swarm and reconfigurable robotics in order to integrate as many trends of modern robotic research as possible.
This paper is organized as follows: In section II we collect the main evolutionary requirements imposed on the platform. In the following three sections III, IV and V we consider the capabilities of autonomous morphogenesis, on-line and onboard evolution and on-board fitness measurement more in detail. The section VI is intended to show the “hardware simulator”. Finally, we conclude this work in Section VII.
The full text can be downloaded from: http://ipvs.informatik.uni-stuttgart.de/BV/symbrion/tiki-download_file.php?fileId=352