Human Gait Trajectory Learning Using Online Gaussian Process for Assistive Lower Limb Exoskeleton

Human gait trajectory estimating and acquiring using human-robot interaction (HRI) is the most crucial issue for an assistive lower limb exoskeleton. The relationship between the HRI and the human gait trajectory is nonlinear, which is difficult to be obtained due to the complex dynamics parameters and physical properties of mechanism and human legs. A Gaussian process (GP) is an excellent algorithm for learning nonlinear approximation, as it is suitable for small-scale dataset. In this paper, an online sparse Gaussian process is proposed to learn the human gait trajectory, i.e., the increment of angular position of knee joints, where the input is the HRI signal and the output is the increment of angular position of knee joints. We collect the HRI signals and the actual angular position by using torque sensors and optical encoders, respectively. When collecting dataset, the subjects are required to wear the exoskeleton without actuation system and walks freely as far as possible. After purifying the dataset, a subspace of training set with appropriate dimensionality is chosen. The subspace will be regarded as the training dataset and is applied in the online sparse GP regression. A position control strategy, i.e., proportion-integration-differentiation (PID), is designed to drive the exoskeleton robot to track the learned human gait trajectory. Finally, an experiment is performed on a subject who walks on the floor wearing the exoskeleton actuated by a hydraulic system at a natural speed. The experiment results show that the proposed algorithm is able to acquire the human gait trajectory by using the physical HRI and the designed control strategy can ensure the exoskeleton system shadow the human gait trajectory.