Passive dynamic walking based on virtual gravity using neural networks

There are two main approaches in order to design biped robots; one is to use precise control and powerful actuators, and the other is based on passive dynamic walking optimizing energy consumption. In this paper a simulation method developed in Matlab® is proposed for a passive robot capable to accomplish a virtual passive walking. The proposed neural network-based method considers three actuators, one at the hip and another two for the ankles. Simulation results are shown for a basic gravity-based control and also for a neural inverse model control.