Notebook Entry

  • Make control contribution plots by taking average of sensors
  • Started a people page on the website
  • Final evaluation for Tarun [September 27]
  • Request quotes from Plone dev to do all this stuff.
  • Got my work computer finally.
  • Brought the gyrobike to the lab.

Walking System Identification

I've added a few things to the analysis and fixed some bugs (see https://github.com/moorepants/DynamicistToolKit/issues?page=1&state=closedfor closed issues.

I have a nicer plot for the model quality comparison which gives the VAF and the RMS of the residuals. This still is validating against the data used to identify the model. My next step is to validate against different data.

I also spent a bunch of time trying to plot the moment contributions due to the error in each sensor and the corresponding gain. To do this, I made the assumption that the commanded sensors during the gait cycle are the mean of the sensors over all cycles. This shows that most of the measured moments are due to the nominal walking moment and the controller contributes to the small deviations away from the nominal motion. If we had perturbed data then the contributions from the controller would be higher and maybe the gains be more certain. These plots didn't really give the kind of insight I was after.

I've limited the sensors and controls to something one may find on a simple exoskeleton design.

The gain plots don't really show any discernible patterns. But removing the left to right side coupling between sensors and controls with the gain omission matrix shows that the quality of the fit doesn't decrease.

The latest IPython notebook with the results is here:

http://nbviewer.ipython.org/urls/raw.github.com/moorepants/walking-sys-id/bd7b6d335d46be04579c085833f76103e8b4ab41/src/control_solver_example.ipynb

Bugs fixed:

  • Signs of Ax and b have been made consistent and correct in the derivation, tests, and source.
  • The estimated controls and plots are now correct.
  • Fixed the variance solution for the gains when a gain omission matrix is present.

Plone

Got a bunch of offers to work on the website. Looks like prices are going to be between 90-165 an hour. The snazzy folks are steeper priced but probably more guaranteed to make something cool. I'm not sure this is worth it though as I can tinker with it for 3-5 times less cost but of course more time. I'll see what some of the other quotes come out to be before deciding.

comments powered by Disqus