An on-line gait generator for bipedal walking robot based on neural networks

An on-line gait synthesis scheme for a bipedal walking robot is proposed. To realize efficient and human-like gait, MTi sensors were mounted on the lower limb of human subject to acquire kinematics information that can be integrated to the angular changes of hip and knee joints during walking. The time series angles were normalized and then sampled by cubic spline interpolation. By employing discrete-time Fourier Series, the samples were extracted into features, and further dimensionally reduced by using PCA to simplified features. By using ANNs, the nonlinear functional relations between gait parameters (i.e. cadence and stride) and simplified features that can be used to reconstruct angles of hip and knee joints were established. Walking experiments of a biped robot at slow, intermediate and fast speeds were conducted to validate the effectiveness of the proposed scheme. The results indicate that the synthesized gait is smooth, efficient and human-like. The proposed scheme can on-line generate the reference gait that covers a wide speed range for bipedal walking of robot.