The following graph shows example joint isolated gains identified from artificial data. I select the marker trajectories and ground reaction loads for a single gait cycle, fit a Fourier series to each measurement, repeat the gait cycles, add Gaussian noise to each marker measurement, and then run it through the data processing pipeline.

This notebook shows the data processing methods:

http://nbviewer.ipython.org/github/csu-hmc/gait-control-direct-id-paper/blob/master/notebooks/marker_error_propogation.ipynb

In this case, it is obvious that something like the vertical GRF is present in the proportional gains. This makes sense, because if you want a model that converts angles and rates to joint torques you'd likely need something that looks like ground reaction loads in your model. But the derivative gains seem like noise centered at zero, which is what you'd hope to get out of this test. I'm not sure what that means.

I'm not sure what do to about this. Should I some how subtract the results from this from the gains I identify from the real data?

This makes me think all we've really developed here is a way to compute the inverse dynamics with a statistical model instead of using multibody dynamics.

@tvdbogert Do you have any thoughts on this?