Reflex-oscillations in evolved single leg neurocontrollers for walking machines

As a prerequisite for developing neural control for walking machines that are able to autonomously navigate through rough terrain, artificial structure evolution is used to generate various single leg controllers. The structure and dynamical properties of the evolved (recurrent) neural networks are then analysed to identify elementary mechanisms of sensor-driven walking behaviour. Based on the biological understanding that legged locomotion implies a highly decentralised and modular control, neuromodules for single, morphological distinct legs of a hexapod walking machine were developed by using a physical simulation. Each of the legs has three degrees of freedom (DOF). The presented results demonstrate how extremely small reflex-oscillators, which inherently rely on the sensorimotor loop and e.g. hysteresis effects, generate effective locomotion. Varying the fitness function by randomly changing the environmental conditions during evolution, neural control mechanisms are identified which allow for robust and adaptive locomotion. Relations to biological findings are discussed.