Authors

Introduction

Manoj's recent paper that got a "control" law.
Wang's predictive simulation
Thomas Geitenbeek predictive simulation
Elliot Rouse's work on the ankle.
van der Kooij paper on direct id
Probably some Kearney on id
Neural network for bicylce tricks from Georgia Tech
Hof's work
GEYER, H., AND HERR, H. 2010. A muscle-reflex model that en- codes principles of legged mechanics produces human walking dynamics and muscle activities
Note the need for high enough perturbations to have any change of getting accurate gain estimates from the controller.
Talk about the control identification problem

Experiments: Basically a citation to the data paper but describe which trials we focus on and why.
Give some plots showing the perturbed vs unperturbed for the trials we had (like the data paper).
Describe the closed loop system architecture.
Desribe the controller: gait phase scheduling, non-linear
Describe direct identification
Describe the linear least squares that provides the results
Describe how we validate the model on independent data
Show that artificially created variations in the data do not produce the same results.
Show what happens if only m* is identified versus all gains.
Give all the input values for the data preparation: filter freq, grf thresholds
Describe the 2D inverse dynamics
Describe the data cleaning

Example gains for one trial
Mean gains for all trials
Show some validation results, i.e. how well the model predictions fit independent data.
Show that it doesn't matter if we compensate the forces/moments wrt belt acceleration.

How do we know if we've perturbed enough?
What does it mean when we get similar results from perturbed and unperturbed data.
Are the forces good quality without compensation?
What are the negative gains all about?
Can we connect this to the physiology?
Do we really have a controller description here?
How can this be useful?