Deep domain adaptation eliminates costly data required for task-agnostic wearable robotic control

Data-driven methods have transformed our ability to assess and respond to human movement with wearable robots, promising real-world rehabilitation and augmentation benefits. However, the proliferation of data-driven methods, with the associated demand for increased personalization and performance, requires vast quantities of high-quality, device-specific data. Procuring these data is often intractable because of resource and personnel costs. We propose a framework that overcomes data scarcity by leveraging simulated sensors from biomechanical models to form a stepping-stone domain through which easily accessible data can be translated into data-limited domains. We developed and optimized a deep domain adaptation network that replaces costly, device-specific, labeled data with open-source datasets and unlabeled exoskeleton data. Using our network, we trained a hip and knee joint moment estimator with performance comparable to a best-case model trained with a complete, device-specific dataset [incurring only an 11 to 20%, 0.019 to 0.028 newton-meters per kilogram (Nm/kg) increase in error for a semisupervised model and 20 to 44%, 0.033 to 0.062 Nm/kg for an unsupervised model]. Our network significantly outperformed counterpart networks without domain adaptation (which incurred errors of 36 to 45% semisupervised and 50 to 60% unsupervised). Deploying our models in the real-time control loop of a hip/knee exoskeleton (N = 8) demonstrated estimator performance similar to offline results while augmenting user performance based on those estimated moments (9.5 to 14.6% metabolic cost reductions compared with no exoskeleton). Our framework enables researchers to train real-time deployable deep learning, task-agnostic models with limited or no access to labeled, device-specific data.