Feature request: per-environment domain randomization of geom contact solver params (sol_params)

[KraHsu]

Hi team — first, thanks for Genesis, it's been great to work with for large-batch RL.

I'm using Genesis for sim-to-real on a contact-rich in-hand manipulation task (a dexterous hand reorienting a cube), and ran into one gap in the per-environment domain-randomization surface that I wanted to raise — and offer to help with.

What already works nicely (per-env DR)

Genesis exposes several per-environment randomization knobs, which is exactly what sim2real needs:

• State-level, always per-env (RigidSolverState, shape (_B, ...)): friction_ratio, mass_shift, i_pos_shift → per-env friction / mass / COM.
• Model-level, opt-in via flags: batch_dofs_info / batch_links_info / batch_joints_info → per-env kp/kv, armature, frictionloss, damping, link inertia, and joint sol_params.

These cover most of the usual DR ranges and work well.

What seems to be missing

Per-env geom contact compliance (sol_params = [timeconst, dampratio, dmin, dmax, width, mid, power], i.e. the MuJoCo solref/solimp analog) does not appear to be randomizable per environment:

• In RigidSolver.set_sol_params(...), the geom branch hardcodes batched = False (only joints_idx/eqs_idx can be batched), so passing a (n_envs, ...) tensor with envs_idx for geoms raises "Expecting at most 1D tensor".
• There is no batch_geoms_info option (only batch_dofs_info / batch_links_info / batch_joints_info).
• RigidSolverState has friction_ratio per geom/env, but no analogous per-env field for sol_params — geom sol_params lives only in the shared geoms_info.

So contact stiffness/damping can only be set globally (shared across the whole batch), whereas friction can already vary per env.

(Version: genesis==1.0.0.)

Why it matters

For contact-rich manipulation, randomizing contact compliance per env (alongside friction) is one of the more important levers for closing the sim2real gap — it's a standard part of DR recipes elsewhere (e.g. MuJoCo solref/solimp randomization). Per-env friction alone covers tangential grip but not normal-direction stiffness/restitution, so the policy can overfit to the simulator's exact contact response.

Possible directions

A couple of options that would mirror existing patterns:

1. Add a per-env sol_params_ratio (or full per-env sol_params) field in geoms_state, mirroring how friction_ratio works, and have the contact constraint kernel read the per-env value — then expose it via a set_geoms_sol_params / envs_idx-aware setter.
2. Or add a batch_geoms_info option (parallel to batch_dofs_info) that makes geoms_info.sol_params per-env.

Is per-env geom sol_params something you'd consider supporting, or is there an existing path I've missed? Happy to help test, and I'd be glad to contribute a PR (option 1 looks like the smaller change since friction_ratio is a good template) if that'd be welcome.

Thanks!

[KraHsu]

Following up after checking the 1.2.0 release: the gap is unchanged. The geom branch of RigidSolver.set_sol_params still hardcodes batched = False (joints and equalities are batched, envs_idx stays dead for geoms), GeomsInfo.sol_params is still allocated as (n_geoms,), and geoms_state still has friction_ratio as the only per-env geom knob.

I also want to walk back option 1 from my original post (the friction_ratio-style per-env ratio). Having read the contact pipeline more carefully, it's the wrong shape for this parameter: a multiplicative ratio doesn't make sense for most of the 7-vector (dmin/dmax/mid are impedances clamped to [IMP_MIN, IMP_MAX], power >= 1, width >= 0), and all validation lives in _sanitize_sol_params at set time (e.g. the timeconst >= 2 * substep_dt clamp), so an in-kernel multiply would bypass it and can feed invalid values into imp_aref. It would also leave two sources of truth next to the existing set_sol_params / get_sol_params.

The codebase already has a pattern for exactly this: batch_joints_info. Applying it field-selectively to geom sol_params behind an opt-in RigidOptions flag would look like:

1. flag in RigidOptions + StaticRigidSimConfig, default False (batch_geoms_info for consistency with the existing trio, or batch_geoms_sol_params if that name shouldn't imply the whole struct is batched — no strong preference);
2. get_geoms_info allocates sol_params with the joints-style shape ternary ((n_geoms, B) vs (n_geoms,)); the other fields stay unbatched since they're topology;
3. kernel_init_geom_fields broadcasts the build-time values through the existing qd.grouped idiom;
4. the geom branch of set_sol_params reads batched = self._options.<flag> instead of the hardcoded False — the public API, its envs_idx parameter, and _sanitize_sol_params (which already reshapes (-1, 7)) are reused as-is;
5. get_sol_params lifts the envs_idx is None assert for geoms when the flag is on;
6. the two mix sites in collider/contact.py (func_add_contact and func_set_contact, the latter covering the backward / split-narrowphase path) index with the existing I_g = [i_g, i_b] if qd.static(flag) else i_g idiom before the 0.5 * (a + b) mean.

With the flag off, layout and kernels are unchanged, so non-DR users pay nothing. Two caveats I'd document rather than solve here: the path stays non-differentiable (ContactData.sol_params carries no grad today), and hibernated contacts keep their pre-mixed values until wake, same as set_global_sol_params now.

If this shape works for you I'll put together a [FEATURE] PR with unit tests (accessor registry rows in test_data_accessor, plus a behavioral check that per-env timeconst changes contact stiffness across envs). @duburcqa would this be acceptable?