Normalized neural network for energy efficient bipedal walking using nonlinear inverted pendulum model

In this paper, we present a novel approach for bipedal walking pattern generation. The proposed method is designed based on 2D inverted pendulum model. All control variables are optimized for an energy efficient gait. To obviate the need of solving non-linear dynamics on-line, a deep neural network is adopted for fast non-linear mapping from desired states to control variables. Normalized dimensionless data is generated to train the neural network, therefore, the trained neural network can be applied to bipedal robots of any size, without any specific modification. The proposed method is later verified through numerical simulations. Simulation results demonstrated that the proposed approach can generate feasible walking motions, and regulate robot's walking velocity successfully. Its disturbance rejection capability was also validated.