Identification of the human postural control system through stochastic trajectory optimization

Background
System identification can be used to obtain a model of the human postural control system from experimental data in which subjects are mechanically perturbed while standing. However, unstable controllers were sometimes found, which obviously do not explain human balance and cannot be applied in control of humanoid robots. Eigenvalue constraints can be used to avoid unstable controllers. However, this method is hard to apply to highly nonlinear systems and large identification datasets.

New method
To address these issues, we perform the system identification with a stochastic system model where process noise is modeled. The parameter identification is performed by simultaneous trajectory optimizations on multiple episodes that have different instances of the process noise.

Results
The stochastic and deterministic identification methods were tested on three types of controllers, including both linear and nonlinear controller architectures. Stochastic identification tracked the experimental data nearly as well as the deterministic identification, while avoiding the unstable controllers that were found with a deterministic system model.

Comparison with existing method
Comparing to eigenvalue constraints, stochastic identification has wider application potentials. Since linearization is not needed in the stochastic identification, it is applicable to highly nonlinear systems, and it can be applied on large data-sets.

Conclusions
Stochastic identification can be used to avoid unstable controllers in human postural control identification.