Online Gait Generation Method Based on Neural Network for Humanoid Robot Fast Walking on Uneven Terrain
Haoran Zhong,Sicheng Xie,Xinyu Li,Liang Gao,Shengyu Lu
DOI: https://doi.org/10.1007/s12555-021-0099-8
2022-03-01
Abstract:Advanced humanoid robots highlight the ability of fast walking and adaptability to uneven terrain. However, owing to the complexity in walking dynamics, disturbances introduced by terrain height variations can adversely affect the bipedal walking performance. Moreover, to generate periodic gaits, most methods require to solve the gait generation problem by using nonlinear optimization approaches, resulting in difficulties for online control. To solve this problem, this paper proposes an online gait generation method to find periodic gaits for fast walking on uneven terrain by using a pre-trained neural network. First, to enhance the terrain adaptability, this paper proposes an improved walking pattern that allows the robots to skip the last single support phase. Such improvement enlarges the feasible step region when stepping down. A compensation strategy is also proposed to reduce the velocity tracking error. Then the improved whale swarm algorithm (IWSA) is applied to generate various datasets that cover the ranges of target velocities and terrain height variations. A back-propagation (BP) network is employed to train these datasets offline to learn the gait dynamics, which is further used to generate the optimal trajectories. Simulation results suggest that, compared with the current methods, the proposed method can solve the walking return map in a short time, with improvements in both maximum walking speed and terrain adaptability.
automation & control systems