Add a time delay term to the controller.

[moorepants]

Something like

m(t) = m*(t) - K1(t) - K2(t - tau)

[moorepants]

Andy Ruina asks this:

In your controller parameters do you have time delay? A shape function for
response vs time (impulse response or step response)? I would think you would
want these. And they might, if you have enough data, vary from sensor to sensor
and joint to joint.

[tvdbogert]

van der Kooij does this, and I think we have enough data to do this also (if we wanted to). The term without delay is due to muscle mechanics and the term with delay is neural. To keep the direct id problem in linear least squares form, we would have to assume a tau rather than estimate it. I would not rule out a nonlinear least squares approach but it's going to be much more computation.

For this paper we should just present the simplest linear controller which lumps everything together into the one term with no delay.

There is the temptation that this may suddenly give great results but we should resist that...