Martin Hülse

Structure and function of evolved neuro-controllers for autonomous robots

The artificial life approach to evolutionary robotics is used as a fundamental framework for the development of a modular neural control of autonomous mobile robots. The applied evolutionary technique is especially designed to grow different neural structures with complex dynamical properties. This is due to a modular neurodynamics approach to cognitive systems, stating that cognitive processes are the result of interacting dynamical neuro-modules. The evolutionary algorithm is described, and a few examples for the versatility of the procedures are given. Besides solutions for standard tasks like exploration, obstacle avoidance and tropism, also the sequential evolution of morphology and control of a biped is demonstrated. A further example describes the co-evolution of different neuro-controllers co-operating to keep a gravitationally driven art-robot in constant rotation.

Robot Control and the Evolution of Modular Neurodynamics

Summary A modular approach to neural behavior control of autonomous robots is presented. It is based on the assumption that complex internal dynamics of recurrent neural networks can efficiently solve complex behavior tasks. For the development of appropriate neural control structures an evolutionary algorithm is introduced, which is able to generate neuromodules with specific functional properties, as well as the connectivity structure for a modular synthesis of such modules. This so called ENS 3 -algorithm does not use genetic coding. It is primarily designed to develop size and connectivity structure of neuro-controllers. But at the same time it optimizes also parameters of individual networks like synaptic weights and bias terms. For demonstration, evolved networks for the control of miniature Khepera robots are presented. The aim is to develop robust controllers in the sense that neuro-controllers evolved in a simulator show comparably good behavior when loaded to a real robot acting in a physical environment. Discussed examples of such controllers generate obstacle avoidance and phototropic behaviors in non-trivial environments.

Synchronization of Internal Neural Rhythms in Multi-Robotic Systems

This paper introduces a method for the coordination of individual action within a group of robots that have to accomplish a common task, gathering energy in a dynamic environment and transferring this energy to a nest. Each individual behavioral pattern is driven by an internal neural rhythm generator exhibiting quasi-periodic oscillations. The paper describes the implementation of this generator, its influence on the dynamics of artificial recurrent neural networks controlling the robots, and the synchronization of internal rhythms with differing frequencies in a group of situated and embodied robots. Synchronization is achieved either by environmental stimuli or even by self-organizing processes solely based on local interactions within a robot population of up to 150 robots. The proposed experimental methodology is used as a bottom-up approach and starting point for answering the question about the complexity required at the individual level to generate sophisticated behavioral patterns at the group level

Dynamical Neural Schmitt Trigger for Robot Control

Structure and function of a small but effective neural network controlling the behavior of an autonomous miniatur robot is analyzed. The controller was developed with the help of an evolutionary algorithm, and it uses recurrent connectivity structure allowing non-trivial dynamical effects. The interplay of three different hysteresis elements leading to a skilled behavior of the robot in challenging environments is explicitly discussed.

The role of non-linearity for evolved multifunctional robot behavior

In this paper the role of non-linear control structures for the development of multifunctional robot behavior in a self-organized way is discussed. This discussion is based on experiments where combinations of two behavioral tasks are incrementally evolved. The evolutionary experiments develop recurrent neural networks of general type in a systematically way. The resulting networks are investigated according to the underlying structure-function relations. These investigations point to necessary properties providing multifunctionality, scalability, and open-ended evolutionary strategies in Evolutionary Robotics.

Co)evolution of (de)centralized neural control for a gravitationally driven machine

Using decentralized control structures for robot control can offer a lot of advantages, such as less complexity, better fault tolerance and more flexibility. In this paper the evolution of recurrent artificial neural networks as centralized and decentralized control architectures will be demonstrated. Both designs will be analyzed concerning their structure-function relations and robustness against lesion experiments. As an application, a gravitationally driven robotic system will be introduced. Its task can be allocated to a cooperative behavior of five subsystems. A co-evolutionary strategy for generating five autonomous agents in parallel will be described.

Evolving Brain Structures for Robot Control

To study the relevance of recurrent neural network structures for the behavior of autonomous agents a series of experiments with miniature robots is performed. A special evolutionary algorithm is used to generate netw orks of different sizes and architectures. Solutions for obstacle a voidance and phototropic behavior are presented. Netw orks are evolved with the help of simulated robots, and the results are validated with the use of physical robots.

Evolving Neural Behaviour Control for Autonomous Robots

An evolutionary algorithm for the creation of recurrent network structures is presented. The aim is to develop neural networks controlling the behaviour of miniature robots. Two different tasks are solved with this approach. For the first, the agents are required to move within an environment without colliding with obstacles. In the second task, the agents are required to move towards a light source. The evolution process is carried out in a simulated environment and individuals with high performance are also tested on a physical environment with the use of Khepera robots.

Representing Robot-Environment Interactions by Dynamical Features of Neuro-controllers

This article presents a method, which enables an autonomous mobile robot to create an internal representation of the external world. The elements of this internal representation are the dynamical features of a neuro-controller and their time regime during the interaction of the robot with its environment. As an examples of this method the behavior of a Khepera robot is studied, which is controlled by a recurrent neural network. This controller has been evolved to solve an obstacle avoidance task. Analytical investigations show that this recurrent controller has four behavior relevant attractors, which can be directly related to the following environmental categories: free space, obstacle left/right, and deadlock situation. Temporal sequences of those attractors, which occur during a run of the robot are used to characterize the robot-environment interaction. To represent the temporal sequences a technique, called macro-action maps, is applied. Experiments indicate that macro-action maps allow to built up more complex environmental categories and enable an autonomous mobile robot to solve navigation tasks.

Modular design of irreducible systems

Strategies of incremental evolution of artificial neural systems have been suggested over the last decade to overcome the scalability problem of evolutionary robotics In this article two methods are introduced that support the evolution of neural couplings and extensions of recurrent neural networks of general type These two methods are applied to combine and extend already evolved behavioral functionality of an autonomous robot in order to compare the structure-function relations of the resulting networks with those of the initial structures The results of these investigations indicate that the emergent dynamics of the resulting networks turn these control structures into irreducible systems We will argue that this leads to several consequences One is, that the scalability problem of evolutionary robotics remains unsolved, no matter which type of incremental evolution is applied.