A Comparison of Four Methods for Predictive Musculoskeletal Simulations of Human Walking

Predictive musculoskeletal simulations are a promising tool for rehabilitation and the design of assistive devices, eliminating the need for extensive data collection. Here, we compare four common methods of predictive gait generation, including two model-based approaches (muscle-reflex and central pattern generator controllers), optimal control and deep reinforcement learning. We also use the same sagittal plane musculoskeletal model, with small method-specific changes, to predict the kinematics, kinetics, ground reaction forces and muscle activations during walking. We validate the results against in-vivo data for healthy walking, and compute root-mean-square errors, Pearson correlation coefficients and the experimental match. Finally, we give examples of model sensitivities and typical deviations of predictive simulations. The results show that model-based methods and optimal control can predict physiological kinematics. However, the latter struggles to predict realistic ankle angles, while the central pattern generator shows significant deviations for the hip and knee. Deep reinforcement learning can be challenging to train, and showed major differences depending on the used musculoskeletal model.