A hexapod walker using a heterarchical architecture for action selection.

摘要: Moving in a cluttered environment with six-legged walking machine that has additional body actuators, therefore controlling 22 DoFs, is not trivial task. Already simple forward on flat plane requires the system to select between different internal states. The orchestration of these states depends velocity and external disturbances. Such disturbances occur continuously, for example due irregular up-and-down movements or slipping legs, even surfaces, particular when negotiating tight curves. number possible further increased allowed walk backward front legs are used as grippers cannot contribute walking. Further necessary expansion allow navigation. Here we demonstrate solution selection sequencing (attractor) required control behaviors at speeds, walking, well negotiation This made by recurrent neural network (RNN) motivation units, bank decentralized memory elements combination feedback through environment. underlying heterarchical architecture allows various combinations elements. modular approach representing an reuse limited procedures adaptation conditions. A way sketched how this may be expanded form cognitive being able plan ahead. characterized types modules arranged layers columns, but complete can also considered holistic showing emergent properties which attributed specific module.