Intelligent path training of a five-link walking robot on sloped surface

Abstract

Intelligent path training of a five-link walking robot on sloped surface is introduced. A neural network theory, backpropagation through time, is applied in this study. The learning scheme uses two neural networks, a neural network controller and a neural network emulator, both of which are multilayered feedforward neural networks. The emulator is trained on accuracy data that characterize the actual walking robot kinematics. The controller learns to provide the control signals at each stage of a walking gait. These trained networks can generate walking patterns by giving reference trajectory which defines the desired step width, height and period in several stages. A mathematical analysis for dynamic walking, based on the ground impact reaction, is included. This proposed scheme is tested with simulations of the BLR-G1 walking robot.