Figure 8.

Algorithm performance as a function of the number of gait cycles trained on. Performance is shown as RMSE (N-m/kg) of algorithm predictions relative to the ground truth values of right and left hip torque in the sagittal plane. For the KF, training data should consist of a minimum of three gait cycles, and little improvement is seen after ten gait cycles. For the CNN, at least 20 gait cycles are necessary for the algorithm to begin improving, and at least 35 gait cycles are needed to reliably outperform the KF. CNN performance continues to improve with additional gait cycles, but only if those gait cycles involve some level of exoskeleton assistance.