Learning a Predictive Model of Human Gait for the Control of a Lower-limb Exoskeleton

Abstract

For an intelligent dynamic motion interaction between a human and a lower-limb exoskeleton, it is necessary to predict the future evolution of the joint gait trajectories and to detect which phase of the gait pattern is currently active. A model of the gait trajectories and of the variations on these trajectories is learned from an example data set. A gait prediction module, based on a statistical latent variable model, is able to predict, in real-time, the future evolution of a joint trajectory, an estimate of the uncertainty on this prediction, the timing along the trajectory and the consistency of the measurements with the learned model. The proposed method is validated using a data set of 54 trials of children walking at three different velocities.

Individualized Gait Generation for Rehabilitation Robots Based on Recurrent Neural Networks

“…RNNs are highly efficient neural networks designed for modeling sequence data such as sentences, voices, and gait patterns. RNNs are naturally more suitable for gait generation tasks than traditional feed-forward neural networks and have been widely used in gait classification [29], [30] and motion forecasting [31], [32].…”

Predicting Seat-Off and Detecting Start-of-Assistance Events for Assisting Sit-to-Stand With an Exoskeleton

“…As described by Aertbeliën and De Schutter [12], Probabilistic Principal Component Analysis (PPCA) [11] can be used to offline learn a time-normalized trajectory f (s) from a dataset of movements:…”

“…In order to estimate the trajectory online, the modes of the offline learned model are estimated using real time measurements. An Iterated Extended Kalman Filter (IEKF) estimates the extended state x = [x * , v] T , which is assumed to be static: h(x k , t) and ∂h ∂x x k ,t are computed as in [12]. The measurement noise covariance R is set equal to σ 2 , the covariance of ε in (1).…”

“…We propose two methods for the prediction of the SO event. The first method is model-based, using a combination of probabilistic principle component analysis [11] and Kalman filtering [12]. The other method is model-free, using a feed-forward artificial neural network.…”

Movement prediction for a lower limb exoskeleton using a conditional restricted Boltzmann machine

“…The nature of the task-predicting both the intended movement type and the future trajectory from time series measurement data-is well suited to methods from machine learning, and not surprisingly many of the existing approaches employ such learning algorithms, whether by using bioelectrical signals or physical measurements, or both. 4,[10][11][12][13] Previous works for the most part adopt a supervisory learning approach, and assume a properly segmented and labeled dataset of trajectories is available for training purposes. In practice, however, it is time-consuming and difficult to measure and collect such trajectories.…”