RoboWalk Trajectory Planning Based on the Human Gait Prediction Using LSTM

Abstract

The increasing demand for rehabilitation robots alongside the complexity of such devices have made their designing procedure a challenging task. One of the challenges is generating a walking pattern for lower limb exoskeletons in such a way that tasks like load carrying or rehabilitation can be performed with minimum disturbance to the human normal gait. This paper presents a deep learning method to predict the next step of RoboWalk joint angle patterns using previous trajectories of former steps. Six different gait data from an individual are chosen as the learning and test dataset. The human joint trajectories during each of the six considered gaits and the augmented human-RoboWalk kinematic model are used to extract RoboWalk’s joint trajectories. The Long-Short-Term Memory (LSTM) network is then used for predicting the future trajectory and classifying the phase of walking. The accuracy for the prediction is about 98.5 percent and the overall error in all the gait modes is less than 5 degrees.