Early Classification of Human Motion Intent for Exoskeleton Assistance Using Kinematics

Exoskeletons can reduce physical effort and the risk of work-related musculoskeletal disorders by providing robotic assistance in repetitive tasks. Designing exoskeletons that seamlessly assist human users without disrupting their natural movements poses a significant challenge, necessitating accurate prediction and adaptation to their motion intent. Particularly, an erroneous detection of motion intent could result in large adversarial effects where the exoskeleton resists the user's desired movement. We propose to analyze the possibility of early and accurate intent detection using only kinematic information and quantify the disturbance that a failed intention detection could entail. We evaluate different classification methods with voluntarily ambiguous experimental data on the intention underlying reaching movements towards four targets in a parasagittal plane. We show that, while recent advances in time series classification -namely using a convolutional and residual neural network-can enable earlier and more accurate intent detection, an informed adaptation of the assistance according to the classification's confidence level is necessary.