Action LQR doesn't pass when I substitute L = np.dot(L.T,L) by L = L.T + L

I noticed that inside the ActionModelLQR we have the following line of code (see https://gepgitlab.laas.fr/loco-3d/crocoddyl/blob/devel/crocoddyl/action.py#L131):

        L = randomOrthonormalMatrix(self.ndx+self.nu)
        L = np.dot(L.T, L)  # ensure symmetric

instead we should have:

        L = randomOrthonormalMatrix(self.ndx+self.nu)
        L = L.T + L  # ensure symmetric

However it's trigger an error inside test_solvers (here https://gepgitlab.laas.fr/loco-3d/crocoddyl/blob/devel/unittest/test_solvers.py#L314) when I changed it. Basically \mathbf{Quu} isn't invertible and we cannot compute the backward-pass.

@nmansard do you understand what it's going on?