mj_setConst

[thowell]

implement mj_setConst
https://github.com/google-deepmind/mujoco/blob/97eb307ce95d67b3a35c8d08bdc7ec0e559a90bf/src/engine/engine_setconst.c#L818

this routine is useful for model editing and should enable safe domain randomization involving certain Model fields on device.
https://mujoco.readthedocs.io/en/latest/programming/simulation.html#mjmodel-changes

• setFixed Add set_const for recomputing model fields after runtime parameter ch… #1086
  • ngravcomp compatible with graph capture
• set0 Add set_const for recomputing model fields after runtime parameter ch… #1086
• setStat Make stat.meaninertia multi-world compatible #1122
• setSpring

[swankystark]

Hi @thowell,

I'd like to work on implementing mj_setConst for mujoco_warp. I've analyzed the C implementation in engine_setconst.c and have a clear understanding of the requirements.

My Approach

I plan to implement this in the following phases:

1. Create mujoco_warp/_src/setconst.py with the main set_const function and helpers:

   • _set_m0: Compute dof_M0 via composite rigid body algorithm
   • _set_fixed: Compute body_subtreemass and ngravcomp
   • _set_0: Compute body_invweight0, dof_invweight0, tendon_invweight0, etc. (dependent on qpos0)
   • _set_stat: Compute model statistics (center, extent, meansize, meanmass, meaninertia)
   • _set_spring: Compute quantities dependent on qpos_spring

2. Leverage existing Warp kernels from forward.py, passive.py, and support.py for kinematics and dynamics computations

3. Focus on runtime-modifiable fields relevant to domain randomization (masses, inertias, statistics), while acknowledging that tree-structure fields are compile-time only

4. Comprehensive testing comparing results against C MuJoCo's mj_setConst to ensure correctness

The implementation will enable safe domain randomization of Model fields on device, as described in the MuJoCo documentation on mjModel changes.

Could you please assign this issue to me? I'm ready to start working on it and can submit a PR within a reasonable timeframe.

Thanks!

[thowell]

@swankystark fyi we have a pr for this feature (or a subcomponent ?) from @kevinzakka #905

@kevinzakka is your pr just for an implementation of _set_0 or does it include the other mujoco routines outlined in the issue description?

probably best to coordinates efforts to prevent duplicate work. thanks!

[kevinzakka]

Hi @thowell, I've updated the PR description, it should explain what's implemented (and what's not).

[thowell]

the pr description is updated with the contributions from #905 (thanks @kevinzakka!)

@swankystark it would be great to contribute _set_stat and _set_spring. for _set_stat we can just compute meaninertia and add todos for the other Statistic fields. does that work for you? thanks!

[therakeshjoshi]

Hi @thowell,
I’d like to work on implementing mj_setConst for mujoco_warp.
I’ve started analyzing the C implementation in engine_setconst.c
and will propose a PR once I have a working version.
Thanks!
Could you please assign this issue to me?

[swankystark]

Thanks @thowell — that works for me.

I’m happy to take ownership of _set_stat and _set_spring as suggested.
I’ll focus on these two components and coordinate with the work in #905 to
avoid any overlap.

I’ll share updates or questions as I make progress. Thanks!

[swankystark]

Hi @thowell, I’ve opened a PR implementing the requested parts of mj_setConst.
This includes set_stat and set_spring, with tests validating against C MuJoCo.
It’s designed to complement #905 and avoid overlap.
Looking forward to your feedback — thanks!