You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I believe that adding differentiability to MJWarp is essential.
Reacted by MartinOpat, Jacob Hackett, Boxi Xia, helloc035, John Viljoen, Davide Risi, abdirahimabdiyassin-cell, Ben Evans, Donghe Chen, Ben Sampaolo and 5 more
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.
Activity
XUZHUOQUN2025 commentedon Jan 31, 2026
I believe that adding differentiability to MJWarp is essential.
MacQWER commentedon Feb 4, 2026
Do you have any plan when the differentiability will be supported?
FirasBDarwish commentedon Feb 11, 2026
It would be great to have differentiability in MJWarp :)
PawitKoch commentedon Feb 17, 2026
Also upvoting for this feature 👍
wzjoriv commentedon Feb 24, 2026
Adding differentiability would be great for new horizons in the ML field. I upvote for the feature 👍
Ruthrash commentedon Apr 1, 2026
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 commentedon Apr 1, 2026
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 commentedon Apr 2, 2026
commenting to show support for this feature :)
chenghuailin commentedon May 13, 2026
+1 for this feature — would love to see it happen soon!
Address review comments
Refresh README, add AGENTS.md, add --view flag to benchmarks/run.py (#…
johnnynunez commentedon Jun 8, 2026
hello guys,
how is it going this?
momo-van commentedon Aug 4, 2026
Please see #1535: hybrid analytic differentiability. We are actively working on this feature request.