Support for differentiability?

[gpu-daal]

Hi all, very excited to see how fast this project is growing!

I was just wondering if Newton supports differentiability. I see that MJ-Warp is the primary backend, and I know it does't support differentiability for now, but what about other backends you guys have? IIRC Warp has some limitations that force you to write kernels in a specific manner for them to be differentiable, which might be a pain.

[preist-nvidia]

@gpu-daal - Newton supports differentiability. You can find some examples here: https://github.com/newton-physics/newton/tree/main/newton/examples/diffsim

[IsshikiHugh]

Hi! I noticed that all the examples there use newton.solvers.SolverSemiImplicit. Does that means the differentiability will only be enabled when this solver is used? Thanks very much!

[momo-van]

Not all solvers are currently differentiable. See this similar thread please #1483

[momo-van]

We will make sure to clearly document this for the next release @eric-heiden for viz.

[preist-nvidia]

Converting this to a discussion.