Towards data-driven predictive control of active upper-body exoskeletons for load carrying

Upper-limb active exoskeletons are a promising technology to reduce musculoskeletal disorders in the context of load-carrying activities. To assist the user on time, it is crucial to predict the assistance torque required for the future intended movement. In this paper, we propose to predict such a torque with predictive models trained on simulated data. We generate exoskeleton sensor data for training learning-based prediction models from human motion capture data. We design a Quadratic Programming control problem for the exoskeleton to track the human body across its movements. From the data generated using this simulation method, we train two torque command prediction methods for transparent control and load carrying. We show that exoskeleton torque command can be predicted with a relative error below 5% at a horizon of 100ms.