crocoddyl
crocoddyl copied to clipboard
loco-3d
Numerical quasi-static torque computation doesn't convergence when we regulariza Cholesky decomposition in contact dynamics
In GitLab by @cmastalli on Sep 24, 2019, 15:25
The contact dynamics is solved using Cholesky decomposition as explained in the following PDF, and details of the code can be seen here: https://gepgitlab.laas.fr/loco-3d/crocoddyl/blob/master/src/multibody/actions/contact-fwddyn.cpp#L64.
The quasi-static torques are computed through a numerical optimization using Newton step, details of the algorithm are here: https://gepgitlab.laas.fr/loco-3d/crocoddyl/blob/master/src/core/action-base.cpp#L30
The problem arises when I used a damping factor in the Cholesky decomposition performed by the forward dynamics. In this case, the algorithm for quasi-static torque computation doesn't convergence.
@nmansard and @jcarpent do you have any clue about this issue?
In GitLab by @cmastalli on Sep 24, 2019, 15:26
changed title from Numerical quasi-static torque computation doesn't convergen{-t-} when we regulariza Cholesky decomposition in contact dynamics to Numerical quasi-static torque computation doesn't convergen{+ce+} when we regulariza Cholesky decomposition in contact dynamics
In GitLab by @jcarpent on Sep 24, 2019, 16:18
I have a solution for it, which will be available soon in Pinocchio. Can you provide some details of your issue on a private discussion?
In GitLab by @cmastalli on Sep 24, 2019, 16:21
I still need this solution since an action model is Pinocchio agnostic. However, I could use the Pinocchio solution for the corresponding cases.
Yes, I will do it.
In GitLab by @jcarpent on Sep 24, 2019, 16:27
This solution will be agnostic too.
In GitLab by @cmastalli on Sep 24, 2019, 16:28
Ok, then let me know when this is available