Control of a passive dynamic compass gait walking robot using neural network central pattern generator
Important features in human walking are the passive dynamic nature of its gait and its rhythmic nature, generated from the Central Pattern Generator (CPG) in a human's spinal cord. The passive dynamic mechanism has been shown in literatures to be a key attribute of the human walking; moreover, passive dynamic walkers, similar to the human body, have been developed that could walk on a downward slope only using the gravity. Much research has been conducted on biped passive dynamic walking mechanisms. Many of these studies focus on either the stability or the control of a walking mechanism. Another feature in human walking is a CPG, a rhythmic movement signal generator. In walking, humans do not plan every step. Instead, a human learns walking patterns and the learned patterns are stored in the human's spinal cord. This attribute has been studied in the past and has been integrated into some biped robots, lately. In this study, a biped model with two features, a passive dynamic mechanism and a central pattern generator, are investigated. The multi-body dynamics equations of motion of biped are formulated and numerically solved. Control of the walking speed of these biped models is the main aim of this study. A model based on Goswami's passive compass gait biped model is used in this study. A CPG-based gait generation method with a feed forward neural network serves as the CPG for the model. The neural network is trained using the acceleration pattern of the legs, rather than the movement pattern of the legs. The selection is based on the human method of learning to walk. Desired walking velocities are selected as commands for the control system. When based on the acceleration pattern of the legs, the control system with the CPG shows robust results in simulations. The results indicate that this CPG-based control system is capable of generating stable gaits and satisfactory results in terms of the walking velocity in a range which is at the outside of the trained data range.
Thesis (Ph.D.)-- Wichita State University, College of Engineering, Dept. of Mechanical Engineering