Contact: Tim Taylor
evolutionary robotics Are robots technical devices that have to be developed and controlled by a human engineer, or could robots also develop and control themselves autonomously?
Traditional robotics that uses Artificial Intelligence planning techniques to program robot behaviors works toward the first, while the Autonomous Robotics approach suggest that the second is a possibility. The robots built according to this approach should be able to adapt to both uncertain and incomplete information in constantly changing environments.
At least two different techniques to reach this goal can be identified. One is to imitate the learning process of a single natural organism. Another, called Evolutionary Robotics, is to reproduce the phylogenetic evolution on populations of robots.
Evolutionary Robotics lets a simulated evolution process develop adaptive robots. By applying selective reproduction on a population of robots the simulated evolution process is directly inspired by the Darwinian evolution theory.
Genetic algorithms can be used to produce control systems for autonomous robots. However, this approach has serious limitations, since it demands considerable time to evolve whole populations of robots.
A simulated/physical approach where main parts of the evolution takes place in a simulator reduces the time consumption dramatically. A gap in performance can however be expected when transferring a robot control system from a simulator to the real environment. This gap in performance can be avoided by letting the robot build the simulator autonomously, adding noise to the simulator, and continuing the evolutionary process in the physical world.
Both the Khepera miniature mobile robot and LEGO robots are used in our Evolutionary Robotics research. The Khepera robot is suitable because of its small size, robustness and accuracy of its PID controllers, and the LEGO robots are suitable especially for co-evolving robot body plans. We believe the evolution of robot body plans to be important, since it is not only the control system of a robot but also the robot body itself that affects the behavior of the robot in a robot system. Experimental results have shown how controllers and robot bodies may co-evolve.
Recent journal publications:
Orazio Miglino, Henrik Hautop Lund, Stefano Nolfi, Evolving Mobile Robots in Simulated and Real Environments. Artificial Life 2:3, 1996.
neuroethological robotics Where Evolutionary Robotics helps us to develop control systems automatically, Neuroethological Robotics can be used to verify the capabilities of known (or pre-defined) control systems. Based on neurophysiological or behavioral evidence, biologists and neuroethologists pose different hypotheses about different animals' control systems. In this context, Neuroethological Robotics provides an empirical field where testing such hypotheses by implementing them on real robots is made possible.
An example is the question on formation of cognitive maps in rats in the open field box experiments. Contrary to the conclusions of biologists, implementations and experiments with a small robot (Khepera) shows that it is not necessary to form a cognitive map to fulfil the task in an open field box: the navigation task can be fulfilled with a simple reactive control system. This tells us that the conclusions about the formation of cognitive maps cannot be based on the open field box experiments.
Another example is our robotics experiments in cricket phonotaxis.
speciation via habitat specialization A long term goal in Artificial Life is to evolve speciation in Artificial Life models. An approach toward this might be to evolve assortative mating, since this is believed to lead to speciation. For instance, sympatric speciation may happen via habitat specialization. Our work on speciation in Artificial Life simulations has been to evolve habitat specialization in populations of neural networks. The emergence of specialization can be due to individual energy extracting abilities, and simulations show that the energy extracting mechanism, the sensory apparatus, and the behavior of organisms may co-evolve and be co-adapted. Populations of organisms have been shown to be pre-adapted to changing environments where the preferred food disappears, and an analysis of the activation of neural network hidden units show that a new food preference can be an ex-aptation, i.e. a new adaptation based on a structure which has previously emerged for adaptively neutral reasons. Further, under social conditions in shared environments it has been found that competition can act to provide population diversification in populations of organisms with individual energy extracting abilities.
Recent journal publications:
Henrik Hautop Lund and Domenico Parisi, Pre-adaptations in Populations of Neural Networks Evolving in a Changing Environment. Artificial Life 2:2 1996.
Other papers Former researcher:
artificial painter Artificial Life techniques can also be used to evolve aesthetic pictures to be used in artistic design. The evolution of pictures is based on the user's aesthetic evaluation of a number of pictures shown on the screen. The Artificial Painter model originates from an inspiration of bio-image techniques such as Computed Tomography (CT), Positron Emission Tomography (PET), and Single Neuron Records. In particular, in AP the techniques of single neuron records are applied on an Artificial Neural Network. The different activation of units in different neural networks gives rise to different, colourful pictures shown to the user on the computer screen. The user can select a number of these pictures to reproduce, and thereby guide the evolution process toward the individual estimate of fitness.
evolution of parallel self-replicating programs An environment is being developed to facilitate the study of patterns of evolution in a system of self-replicating entities which are competing with each other for resources. In this environment, computer programs are the self-replicators, and large numbers of them compete with each other for the memory and CPU time required to make copies of themselves. The programs are also subject to mutation, so that, over time, mutants which are better at making copies of themselves become more numerous in the population of programs, and a process of evolution is observed.
The programs in this system may be considered as being somewhat analogous to the unicellular biological organisms that were around in the Vendian period of geological time, immediately prior to the emergence in the fossil record of macroscopic multicellular organisms. The instructions that make up a program are its `genotype', which gets interpretted into the action or behaviour of the program (its `phenotype').
The environment is called COSMOS (standing for COmpetitive Self-replicating Multicellular Organisms in Software). It bears some similarity to Tom Ray's Tierra system, but COSMOS is designed specifically to look at the evolution of parallel programs (the biological analogy being multicellular organisms).
COSMOS uses a distributed memory, MIMD model of parallelism (in contrast to the work that had been done on parallel programs in Tierra). This being the case, it has to rely heavily on communication between processors to perform parallel computations. In fact, message passing is the only way that one process (`cell') within a multi-process (`multicellular') program (`organism') can communicate with its neighbours, or with other programs. Apart from this inter-cellular communication, each cell only has read and write access within its own cell boundary.
There are also a number of other fairly important differences between COSMOS and Tierra, designed to encourage the evolution of diversity and complexity in the competing programs, rather than just the optimization of their ancestral algorithms.
links to other ALife-related pages