A novel soft computing architecture for the control of autonomous walking robots

An integration of concepts from neurobiology, applied psychology, insect physiology and behaviour based robotics has led us to propose a novel generic systems architecture for the intelligent control of mobile robots and in particular, autonomous walking machines. (We define what we mean by “autonomy”.) The control architecture is hierarchical and will be described from a top-down perspective. Level one consists of interpreting a motivation and translating this into high-level commands. Once a high-level command is generated, a range of internal representations or “cognitive maps” may be employed at level two to help provide body-centred motion. At level three of the hierarchy kinematic planning is performed. The fourth level – dynamic compensation – requires feedback from the actuators and compensates for errors in the target vectors provided by the kinematic level and caused by systematic dynamic uncertainties or environmental disturbances. This is implemented using adaptive neural controllers. The interfaces will be described and results from simulation and implementation of levels 2–4 on a hexapod robot will be presented. The hierarchy employs the following soft computing techniques: evolution strategies, cognitive maps, adaptive heuristic critics, temporal difference learning and adaptive neural control using linear-equivalent neural networks.