Static Optimisation using muscle equilibrium #2404
Conversation
pariterre
changed the title
pariterre
changed the title
|
Up? |
|
@aseth1 Can you comment on whether this functionality needs to be supported by mainstream OpenSim rather than as a plugin? |
|
@aymanhab The idea is to linearize at the point of previous activation instead of full activation. My implementation doesn't fully mimic what happens if you don't linearize at all (that is commenting the line that defines USE_LINEAR_CONSTRAINT_MATRIX), but is way closer than the actual linearization. For instance, for the arm26, the actual implementation shows differences up to 30% while this new linearization method show differences up to 2%. As an outside project, I've also coded a way to use the non-linearize method by approaching the constraint jacobian by series of cublic spline. It therefore constructs the Jacobian at the beginning of the optimization and then approximates it using the derivatives of the spline. Using this method, I get almost no difference with the non-linearized method. However, it takes about twice the time than the linearized method (which is still almost 500X faster than original non-linear method). I did not planned to PR this method to OpenSim though, unless you guys think that it would make sense to do so.. Thanks! |
|
Up? |
|
@aymanhab |
|
@jimmyDunne I'm happy to help review this after discussing with the dev team what we want to do. |
|
@pariterre Thank you for making this (and other) pull requests! It is great to have contributions from the community. Since this is a pretty large addition to the code, could we set-up a time to talk via Skype about how best to incorporate this work? You can email us at opensim@stanford.edu and we can arrange a time to chat. Thanks again! |
| ScalarActuator* act = dynamic_cast<ScalarActuator*>(&fs.get(i)); | ||
| if (act) { | ||
| if (act->getMinControl() != -INFINITY) | ||
| _parameters[j++] = act->getMaxControl(); |
There was a problem hiding this comment.
Do you mean act ->getMinControl()?
There was a problem hiding this comment.
Actually, the error is in the if. Unfortunately, contrary to what is previously done, the initial guess on the first frame has to be maxControl. The reason is that I linearize between 0 (actually MinControl) and value from previous frame. If we set the previous frame to be minimal, it is a singularity. Setting it to max solves to the same value as the previous linear version : the problem being quadratic when p=2 and therefore the solver finds a good optimal solution even though I initialize at max.
|
Up? :) |
…ed) and linearized is now compliant to muscle equilibrium
|
Up? @aymanhab |
Fixes issue #500
Brief summary of changes
Static optimisation (without linearization) now solves for muscle equilibrium when calculating the acceleration. So instead of solving the "realize acceleration" via a an infinite tendon stiffness assumption, it now solves for the proper equilibrium.
Testing I've completed
None, this breaks the previous values that were considered as reference.
In that sense, I will need your opinion on what I've done.
CHANGELOG.md (choose one)
This change is