Gait Modeling And Learning For Lower Limb Rehabilitation Exoskeleton Robots

Generating natural and safe gait patterns is an critical issue in the study of lower limb exoskeleton robots for patients with gait disorders.Due to the complex terrains in our daily life and the anthropometric variations for different patients,there are two challenges for gait planning of lower limb exoskeleton robots:terrain adaptive gait planning and individualized gait planning.In this dissertation,different gait models and learning algorithms were proposed for the terrain adaptive gait planning,the individualized gait planning for paraplegic patients,and the synergetic gait prediction and assist-as-needed control strategies for hemiplegic patients.Main contributions can be summarized as:For the terrain adaptive gait planning,this dissertation proposed a linear inverted pendulum model and dynamic movement primitives based adaptive gait planning approach for walking assistance lower limb exoskeleton robots.The gradient of the sloped terrain can be estimated with the sensor data from the exoskeleton robot,the human-exoskeleton system can be modeled based on the linear inverted pendulum model,and an adaptive gait planning approach based on dynamic movement primitives can be constructed.As a result,the gait models can be learned from the the gait database of healthy subjects,and generate adaptive gait trajectories for slopes with gradient from 0° to 12°.For the individualized gait planning of paraplegic patients,this dissertation proposed an individualized gait prediction model consists of a gait parameters model and a gait trajectory model.The gait parameters model is constructed with neural networks,and the gait trajectory model is constructed with the kernelized movement primitives.The proposed gait models can be learned from the gait database of healthy subjects and generate personalized gait trajectories according to the desired walking speed and the patient's anthropometric parameters,which is adaptive to different patients with heights from 155 cm to 185 cm as well as the desired walking speed from 0.5 km/h to 6.0 km/h.For the synergetic gait prediction and assist-as-needed control for hemiplegic patients,a synergetic gait prediction model based on the recurrent neural networks and attention mechanism as well as assist-as-needed control approach based on path control strategy are proposed in this dissertation.The synergetic gait prediction model can predict the future synergetic gait trajectories based on the patient's historical joint angles,which is adaptive to the walking speed from 0 km to 6 km/h.The assist-as-needed control approach provides the required assistance joint torques according to the patient's own motor ability,enhancing the active participation of both the healthy and the paretic side for hemiplegic patients in gait training.The proposed gait models and control approach allow the patient to actively guide the training process through the healthy leg and control gait parameters such as stride length and frequency,making the training process more personalized for the patient.The proposed gait models and learning approaches were validated on the walking assistance exoskeleton robot named AIDER and the rehabilitation exoskeleton robot named HemiGo,which provide possible gait modeling and learning solutions for the exoskeleton robot in terrain adaptive gait planning and individualized gait planning,and can be references for the research of the adaptive gait planning for exoskeleton robots in future.