Markus Gambietz (Erlangen), Marlies Nitschke (Erlangen), Dr. Jörg Miehling (Erlangen), Prof. Anne D. Koelewijn (Erlangen)
Abstract-Text (inkl. Referenzen und Bildunterschriften)
Introduction
Simulation studies have previously identified movement objectives that represent human gait in optimal control simulations [1,2]. However, they had to employ speed constraints. Therefore, previous simulations could not predict the effect of interventions, such as prostheses, on walking speed and cadence simultaneously. Thus, we aimed to identify objective weights (OW) to create optimal control gait simulations without predefining spatiotemporal parameters.
Methods
We applied bi-level inverse optimal control (IOC) on gait cycles (GCs) with two different speeds [3]. Our goal was to find OW that minimized the difference between the resulting predictive 2D musculoskeletal optimal control simulations and the measured optical motion capture and force plate data. We optimized the OW for six movement objectives from [1,4] using Tree-Parzen Estimators [5].
Based on the weights resulting from IOC, we fitted a linear regression model that estimated OW from walking speed and GC duration. We solved predictive simulations of 48 grid-sampled combinations of speed and GC duration using the estimation from the fitted model and evaluated whether the desired speed and GC duration matched.
Results & Conclusion
Fig. 1 shows that the GC duration matched well, while speed was often too low. Furthermore, 13 simulated GCs did not yield walking behavior, since for some the speed/duration was sampled out of distribution, while for others an unrealistic local minimum was found. In conclusion, our proposed approach did successfully create unconstrained gait simulations. However, accurate speed representation needs to be improved.
References
[1] Falisse et al., J R Soc Interface (2019)
[2] Koelewijn et al., Comput Methods Biomech Biomed Engin (2018)
[3] Mombaur et al., Auton Robots (2010)
[4] Veerkamp et al., J Biomech (2021)
[5] Akiba et al., KDD (2019)
Fig. 1: Comparison of speed and GC duration between linear regression inputs and optimal control simulation results.