tree 13c06cf26044d7ae8ae70b0e6ff539e0fb47b78b
parent 194511c9bfa26dc5b4027c7180e9554f2ce83e65
author Sergiu Deitsch <sergiu.deitsch@gmail.com> 1791072610 +0200
committer Sergiu Deitsch <sergiu.deitsch@gmail.com> 1791191621 +0200
gpgsig -----BEGIN SSH SIGNATURE-----
 U1NIU0lHAAAAAQAAADMAAAALc3NoLWVkMjU1MTkAAAAg3vg8qwvHVhPBrnx4YDqFxqFP3d
 j6luLZ1zAM1x55LLkAAAADZ2l0AAAAAAAAAAZzaGE1MTIAAABTAAAAC3NzaC1lZDI1NTE5
 AAAAQKjJ38Wk2pdhANGbY+L1in/J/ePW93YDDG8uO73FgE5onDqr3XXz8ARwye3pCnU/pl
 SQmJa6Nsv5fa0xMpQXfA8=
 -----END SSH SIGNATURE-----

Do not differentiate the norm at gimbal lock

At gimbal lock, the norm determining the middle Euler angle is exactly
zero. Its kink carries over through atan2, so the angle has no
derivative. Differentiating the norm anyway produces NaN derivatives.

Compute the norm only away from gimbal lock and use its zero subgradient
otherwise. Likewise, compute the norms in the quaternion manifold test
functor only if they are nonzero.

Change-Id: Ic5c4eaaf1b268b7f002c1489296b212fd45f7727