Simplifying neural networks for controlling walking by exploiting physical properties

Abstract

A network for controlling a six-legged, insect-like walking system is proposed. The network contains internal recurrent connections, but important recurrent connections utilize the loop through the environment. This approach leads to a subnet for controlling the three joints of a leg during its swing which is arguably the simplest possible solution. The task for the stance subnet appears more difficult because the movements of a larger and varying number of joints (9–18: three for each leg in stance) have to be controlled such that each leg contributes efficiently to support and propulsion and legs do not work at cross purposes. Already inherently non-linear, four factors further complicate this task: 1) the combination of legs in stance varies continuously, 2) during curve walking, legs must move at different speeds, 3) on compliant substrates, the speed of the individual leg may vary unpredicatably, and 4) the geometry of the system may vary through growth and injury or due to non-rigid suspension of the joints. We show that an extremely decentralized, simple network copes with all these problems by exploiting the physical properties of the system.