Can differentiable performance be applied to the training of the anymal_d quadruped robot?

[Mirrency]

I would like to inquire whether Newton's differentiability can be used in the training of the anymal_d quadruped robot? I noticed that in the diffsim examples, only newton.solvers.SolverSemiImplicit is used, while the robot uses newton.solvers.SolverMuJoCo. MuJoCo does not seem to fully support differentiability internally. When I apply newton.solvers.SolverSemiImplicit to anymal_d, there may be a situation of state explosion, causing me not to see the loaded model.

[momo-van]

@Mirrency MuJoCo Warp is not currently differentiable. This will be documented clearly here #904

Please see similar topics here:
#1762
#1483