Robust Human Movement Prediction by Completion-Generative Adversarial Networks with Huber Loss

In recent years, wearable exoskeleton robots have been growingly used for rehabilitation or movement assistive purposes. Despite the growing application of these robots in various domains, such as physical therapy, the movement synchronization between robots and human bodies remains a challenging problem. This paper aims to achieve better synchronization by predicting human movement. Although several works have been presented in this domain, the robustness of these predictions has received less attention. This paper aims to provide a robust prediction using Completion-Generative Adversarial Networks (CGAN) that are learned based on the Huber loss function. Specifically, we reshape the 3D-joint-position-time series (jointxaxesxtime) into multivariate time series ((jointxaxes) xtime) and pass them to a CGAN. We use the Huber loss function to improve the GAN performance and offer higher robustness against noise in real-world applications. The proposed method is evaluated on an actual human gait dataset and compared with several recent works in this domain. Results show that the proposed method is superior to the previous works in prediction error, particularly in terms of achieving a better signal-to-noise ratio.