[Question] Does Isaac Lab 3.0 (latest dev branch) support differentiable physics for BPTT?

[hddbuaa]

Hi Isaac Lab team,

I am currently working on Differentiable Reinforcement Learning (specifically algorithms similar to BPTT - Backpropagation Through Time) and testing it within the latest development branch of Isaac Lab 3.0. My setup relies on the Newton physics engine to simulate UAV (equpped with wheels and arms) dynamics.

I have already implemented a version of the training code. However, I've run into an issue during the backward pass: the state variables returned by the simulation environment have no gradients attached (the computation graph seems to be detached after the physics step).

I would like to clarify a couple of things regarding the current capabilities:

Does the current latest main/dev branch of Isaac Lab 3.0 (specifically with the Newton backend) actually support end-to-end differentiable physics simulation?

If differentiable physics is not currently supported, is there an estimated roadmap or timeline for when this feature might be available for training?

Any insights, workarounds, or pointers to relevant branches/documentation would be highly appreciated.

Thanks in advance for your time and the great work on Isaac Lab!