[QUESTION] How to use Autodifferentiation for soft-body parameters

[rrzhang139]

Hello, thank you for building this first of all.

I have a question about differentiating the parameters in Warp. If I wanted to differentiate soft body parameters like k_mu or k_lambda in soft_grid, how would that generally work? Would I initialize some tensors and pass them into kernels to get that working? Do you have any code samples that does this for soft body simulation? Thanks!

[amabilee]

Hey there !

To differentiate soft body parameters like k_mu or k_lambda in Warp, you indeed need to initialize tensors and pass them into kernels. Warp supports automatic differentiation, which allows you to compute gradients of these parameters.

Here is the doc about it:
https://nvidia.github.io/warp/modules/differentiability.html

[shi-eric]

I'm not an expert on the warp.sim things, but I'll attempt an answer. You would generally need to do two things:

• When calling the finalize() method of the ModelBuilder class, you need to use requires_grad=True to ensure that gradients are calculated for the Model object arrays.
• After recording the forward pass onto a wp.Tape() and calling the backward() method with the appropriate loss input, you would then get the gradients of the Model.tet_materials array. The details of this array are here: https://nvidia.github.io/warp/modules/sim.html#warp.sim.Model.tet_materials