(EN)Assistive soft robotic exosuits are a promising avenue for enhancing independent living for individuals with mobility restrictions. Wearable sensors can be integrated into these exosuits, which can provide data for activity recognition. Such data can be used to train Machine Learning models for movement prediction in order to provide real-time assistance via actuation of a soft exosuit.

Models such as these must be highly accurate for activity classification (e.g. going from seated to standing). They must also make predictions using small time windows of data on low latency edge devices. This minimises the time between user movement and the classification of an activity to enable real-time assistance.

For our models to be trained on diverse datasets of multiple participants whilst maintaining user privacy we must use Federated Learning, where models learn from distributed data. Models must also remain well-calibrated to individuals who might vary in movement quality over time (e.g. as medical conditions progress, or injuries are rehabilitated).