CNN factored sEMG based Limb Angle Prediction on the Edge - Advancing Model Predictive Control of Robotic Rehabilitative Systems

Robotic rehabilitative systems promise to revolutionise the future of stroke rehabilitation. Controlling motors to allow the patient to move at their will during parts of the motion that they don’t require assistance is important for effective rehabilitation, especially at the later stages of therapy. However contemporary rehabilitative systems use a force-based feedback system that provides a small resistance to the motion as the system adapts only after the interaction between the patient and the machine. However, if the motion is predicted in advance, and further engineered in the rehabilitative system, these small resistive experiences are unfelt. Considering the critical method of usage and vulnerable segment of users, prediction on the edge to satisfy timing becomes imperative, besides running the process on constrained resources is expected. As a step towards this model predictive control approach, this paper proposes a limb angle prediction strategy based on surface Electromyography (sEMG) signals. The proposed strategy that is established from Convolutional Neural Network (CNN) is demonstrated and validated on the edge for upper limb bicep flexion and extension movement. The paper measures the impact of various design parameters on the implementation of the system and its relations to prediction accuracy, prediction time and resource requirement on the edge. Dataset comprising of sensory systems employed, and prediction models are made freely available for easy adoption and further usage to the researchers and designers community.