A Lightweight Cable-driven Hip Power Assisting Exoskeleton utilizing VAE-CTNet based Hip Trajectory Prediction Control

Achieving efficient human-robot interaction in lower limb power-assisted exoskeletons (LLPaE) critically relies on accurate trajectory prediction (TP)-based control. Nevertheless, conventional data-driven TP methods typically necessitate large datasets and exhibit limited generalization performance in data-scarce scenarios. In this study, we developed a lightweight cable-driven LLPaE named Nu-HPaE, which provides hip assistance via motor-driven cables, and utilizes accurate TP to realize adaptive control. A novel TP model called VAE-CTNet, which integrates a Variational Autoencoder (VAE), Convolutional Neural Network (CNN), and Transformer architecture is proposed. VAE-CTNet achieves accurate TP by data augmentation through the VAE and capturing multi-scale temporal local features and long-range dependencies within motion sequences using the CNN and Transformer. A self-collected dataset encompassing six common daily movements using Inertial Measurement Units (IMUs) is applied to VAE-CTNet training. Experimental analysis demonstrates that the integrated VAE module effectively performs data augmentation by generating diverse and physiologically plausible hip joint trajectories, significantly enhancing generalization. Furthermore, in Nu-HPaE prototype tests, four motion sequences are designed as combinations of the six daily motions, validating VAE-CTNet’s real-time TP performance as well. Meanwhile, comparative analysis of surface electromyography (sEMG) signals from the rectus and biceps further confirms the power assisting effects provided by Nu-HPaE. The proposed VAE-CTNet achieves high-precision TP in Nu-HPaE, validating its significant potential for adaptive human-robot collaboration. Also, the power assisting effects provided by Nu-HPaE validates the rationality of the exoskeleton’s design.