Differentiability

[thowell]

https://nvidia.github.io/warp/modules/differentiability.html#differentiability

[XUZHUOQUN2025]

I believe that adding differentiability to MJWarp is essential.

[MacQWER]

Do you have any plan when the differentiability will be supported?

[FirasBDarwish]

It would be great to have differentiability in MJWarp :)

[PawitKoch]

Also upvoting for this feature 👍

[wzjoriv]

Adding differentiability would be great for new horizons in the ML field. I upvote for the feature 👍

[Ruthrash]

A differentiable renderer would allow system Identification workflows that estimates system params by by minimizing rendered and real images. Right now this sysID process is done manually by visually comparing a video in real and in sim
https://isaac-sim.github.io/IsaacLab/main/source/policy_deployment/02_gear_assembly/gear_assembly_policy.html#:~:text=These%20friction%20values%20(0.75)%20were%20determined%20through%20iterative%20visual%20comparison%3A

[wzjoriv]

A differentiable simulator would allow one to more easily develop MPC, DDP and ILQR controllers using MuJoCo model. That means trajectory planner and optimal control for digital twins defined in mujoco.

While JAX MJX already supports this via mj_forward, if Warp can be a superior alternative with more efficient performance it would be great. These methods often require batched computation of Jacobian and Hessian with respect to states and controls along a trajectory. Specially, efficient second-order derivatives would be a significant value-add over existing XLA-based implementations.

It is great to see a PR #1226 looking into it.

[johnviljoen]

commenting to show support for this feature :)