How to use differentiable simulation?

[HerocatUED]

After we set requires_grad=True in SimOptions, how can we get gradient from gs scene or env?
Say, we have action_t = model.act(state_t) and state_{t+1} = env.step(action_t), our loss function is L2 norm between state_{t+1} and target_{t+1}, how can we calculate gradient to optimize our model or policy?

[HerocatUED]

Oh, I forgot to point out that the env above is for a unitree-Go2 locomotion task. Additionally, I don't know which method should I use for rigid body simulation. I just use default settings as example does. When I check API doc, I noticed that some solver is not differentiable currently. And I know nothing about physics solvers. So, I wonder will there be any examples for differentiable simulation.