Inquiry regarding future support for Newton

[RHnejad]

Hello,

I was wondering whether there are any active development plans to support MyoSuite on top of [Newton]?

If not, I would appreciate your insight on:

How challenging would it be to implement a custom port?

What are the primary technical hurdles (e.g., muscle dynamics, API differences) to look out for?

Thank you for your time!

[Vittorio-Caggiano]

Hi @RHnejad,
we have not discussed Newton support but we have been testing mujoco-mjx / mujoco-warp for hardware acceleration. It is currently staged here for test/preview and it will be part of next major myosuite release.

We have also been testing mjlab https://github.com/MyoHub/mjlab_myosuite for further extensions and compatibility.

What do you think would be the major advantage to have Newron as opposed to mjlab or mujoco_warp native implementation?

[RHnejad]

Thank you so much for the reply and the update! It’s nice to see mjx and warp are already being tested.

The main reason I was thinking about Newton specifically is the differentiability aspect. I was wondering if we could get better guidance for muscle signals through Newton's warp kernels compared to standard autodiff in jax/mjx, especially since musculoskeletal dynamics can get pretty 'stiff' and hard to optimize.

Also, as far as I understand, Nvidia deprecated warp.sim and moved that development over to Newton. I’m a bit worried that building a custom mujoco-warp implementation might hit a dead end later if the ecosystem moves fully toward Newton as the standard.

P.S. I am actually pretty new to myosuite, so I am still getting my bearings, but I wanted to make sure I'm starting on the most future-proof foundation. I am already moving my other non-musculoskeletal projects over to Newton because it shows so much promise for future research in robotics.