Prof. Frank Pasemann (Wissenschaftskolleg zu Berlin & Uni Osnabrück): "Evolutionary Robotics, Neurodynamics, and Behaviour Control of Animats"
Evolutionary Robotics is a research field of intrinsically interdisciplinary character. It takes up approaches from and contributes to fields like neuroscience, cognitive science, computer science, and robotics. Using evolutionary computation techniques the objectives of Evolutionary Robotics include, among others, the development of neural behaviour control for physical and/or simulated robots. It aims also to contribute to a deeper understanding of (behaviour relevant) neural processing in biological nervous systems by analysing evolved artificial ones.
Working in the context of embodied and situated cognition, here it is assumed that cognitive abilities are based on the dynamical properties of the underlying biological or artificial nervous systems. Reviewing shortly the dynamics of recurrent neural networks, a description of the Artificial Life approach to Evolutionary Robotics is given. And the role of appropriate evolutionary environments is discussed.. Some typical results for neural behaviour control of autonomous robots (animats) are presented. Examples include wheel-driven robots as well as walking machines and humanoids.
The motor pattern in forward walking stick insects is generated by the interplay of central rhythm generators (CRG) and sensory feedback from the legs. This sensory feedback enables the insect to adapt its walking and climbing movements to variations in substrate conditions on a step by step basis. Coordination of the legs is achieved by state-dependent influences of the adjacent neighboring legs. In stick insects different walking directions can be elicited by specific tactile stimulations. Touching the abdomen induces forward walks whereas backward walks can be in-duced by gently pulling the antennae. Both situations require distinctly different motor patterns, as e.g. the functional swing and stance muscles are inverted. It has to be mentioned that continuous backward walking – a behavior not expressed spontaneously by the animal – appears to be less regular than forward walking. Thus, backward walking might be an emergency program that might be produced predominantly by the actions of CRGs and less precisely controlled by sensory feedback. Thus, I investigated the starting behavior of forward and backward walking and how the motor patterns of the middle leg are initiated and concerted and which role can be attributed to sensory feedback of the leg.
Several neural subsystems interact in the brain of the bee during tactile learning. Sensory stimuli evoke specific antennal motor programs that optimize the sampling of sensory information. The dendritic projections of gustatory antennal sensilla into the antennal motor neuropil are the neural substrate for interactions of gustatory information with the motor system.
Motor learning is the simplest form of associative plasticity at the level of the antennal motor system. Operant conditioning of antennal activity with sucrose demonstrates that the association of the salient gustatory stimulus with motor activity can induce rapid and long-lasting modifications of antennal motor patterns. Operant conditioning can be reduced to a single motoneuron. Tactile PER conditioning represents the highest level of plasticity in this system. In this conditioning protocol bees learn the position of an object and they learn to discriminate form, size and surface textures. Specific antennal scanning patterns are a necessary prerequisite for the discrimination of surface textures.
Honey bees that perform different foraging tasks (nectar-, pollen-, water-collectors) differ in their sensory thresholds for different stimulus modalities. These differences in sensory thresholds are dependent on a number of intrinsic and extrinsic factors and they could provide a basis for the division of labor within a colony. Differences in sensory thresholds result in differences in learning performance. Experimental data suggest that honey bees can learn the characteristics of different resources because sensory thresholds are adapted to the foraging conditions in the field.