Our work on robust thought-controlled exoskeletons in collaboration with the Utah Bionic Engineering Lab is now published online in Frontiers in NeuroRobotics. We explored how a powered hip exoskeleton impacts muscle activity, and the implications of that on real-time EMG control. We showed that lower-limb and lower-back muscle activity change non-linearly as a result of increasing exoskeleton assistance - this makes real-time EMG control difficult because the act of controlling (assisting) changes the input signal. The good news is that nonlinear neural networks are capable of generalizing predictions of torque across different levels of exoskeleton assistance, when explicit training data is provided. However, a common linear model (i.e., a Kalman filter) is not capable of the same generalization. Lastly, given that explicit training data on every level of exoskeleton assistance may not be feasible, we show that, when time is limited, training data for EMG control of exoskeletons should use at least 35 gait cycles and emphasize the highest levels of assistance first! The full article is available open-access here: https://doi.org/10.3389/fnbot.2021.700823
