Getting the gradients of joint targets using the XPBD integrator

[Manchewable]

Hi,

I'm wondering whether it's possible to compute the gradients w.r.t. model.joint_target using the XPBD integrator.

Whenever I tried to get its gradients via tape.backward(), I'm getting a zero gradient in return.

Appreciate to any insights!

[eric-heiden]

Hi @Manchewable,

We are currently in the process of improving the differentiability of the XPBD integrator. Differentiating w.r.t. joint targets will be possible in a future version. For now, you could try to ensure that you allocate a new state for every substep of the trajectory that is rolled out (so the states are not overwritten but instead state_out is mapped to a new state when you call integrator.simulate()), and that the Warp arrays in the state and model have requires_grad=True defined.