Scale operation uses compiled defaults for scalar material fields.

Values = default * sample, within expected bounds.

SceneEntityCfg in params should not be mutated when resolved.

Test that scene initialization sets up entities.

Base scene sets needs_update when contact overlays are visible.

Test that reset calls reset on all entities.

Test that spec_fn can add elements to the scene spec.

Test that spec_fn can create tendons between two entities.

Test that update calls update on all entities.

Test creating a scene with multiple entities.

Test that write_data_to_sim calls the method on all entities.

Scene.write produces a directory; with zip=True, a zip archive.

Per-world defaults are indexed by env first, then by entity.

Selects all GPUs when 'all' is specified.

Selects CPU mode when CUDA_VISIBLE_DEVICES is empty.

Selects CPU mode when None is specified.

Handles empty segments in CUDA_VISIBLE_DEVICES.

Handles MIG GPU UUIDs in CUDA_VISIBLE_DEVICES.

Preserves the order specified by the user.

Correctly selects a single GPU.

Handles non-contiguous CUDA_VISIBLE_DEVICES correctly.

Handles CUDA_VISIBLE_DEVICES with spaces correctly.

Respects CUDA_VISIBLE_DEVICES when selecting GPUs.

Verify sense_graph is captured on CUDA.

Sensor and entity both equal hand-computed projected gravity for a tilted base.

VariantEntityCfg.spec_fn is unused; setting it should fail loudly.

Verify that per_env=False gives same lag across all envs.

All entities within same env get same value, envs differ.

Each group owns its own noise model instance when they share a term name. ConstantNoiseCfg(op="add") shifts the additive-bias tensor by the configu

Test that sim.reset() restores qpos/qvel to initial values.

Test that sim.reset() only affects specified environments.

Test that SimulationCfg values are applied to both mj_model and wp_model.

Test that a single entity produces one merged keyframe.

num_envs=1 works without expand_model_fields, no accumulation.

Test observation with history_length=1.

A single variant degenerates cleanly; useful for templated variant sets.

ray_alignment only controls direction, not origin position. When a calf body pitches, the foot site swings to its physical position (site_xpos).

The full DR path: running ``dr.site_quat`` perturbs the gravity observation. This is what the G1 example configs rely on -- randomizing the IMU sit

``slew_rate`` rate-limits ``ctrl``: after one step, effective voltage is far below the requested input.

variant_slot_specs[v] aligns with slots positionally for every variant.

Articulated variants produce slots per (body, role).

A geom with contype=0 and conaffinity=0 is classified as visual.

Within a (body, role), slot ordinals are 0..max-1.

A slot held by some variants but not others is unfilled (None) for those.

source_geom_names has None where a variant doesn't fill a slot.

Every template slot name starts with the reserved mjlab/pad/ prefix.

Slot ordering: visual slots come before collision slots per body.

Visual and collision geoms on the same body get distinct slots.

A tiny but planar patch should not be rejected as degenerate.

All-limited joints should have finite soft limits.

All-unlimited joints should have infinite soft limits.

Unlimited joints must have [-inf, inf] soft limits, not [0, 0].

sort_actuators=True reorders ctrl to match joint definition order.

Verify spatial frame sensors accept body, xbody, geom, site, camera.

Verify that without update(), cache returns stale data.

Test that stateless class-based rewards without reset don't break reset.

Step-mode events fire unconditionally on every apply() call.

VideoRecorder writes a readable mp4 when the step trigger fires.

Physics steps match sim-time budget.

Average stone size should be smaller at higher difficulty.

Per-component DR with regex patterns.

Pattern matching no names raises ValueError.

Stripping buffer textures removes them and enables XML roundtrip.

Reset zeroes substep accumulators for the given env_ids.

Test contact detection with subtree mode.

Viewer model sync copies explicit fields and leaves others unchanged.

Entity.reset() zeros all targets.

Template contype/conaffinity are derived from slot role (union of variants).

All entity mesh-geom names in the compiled template use mjlab/pad/ prefix.

Test that tendon/site targets are only allocated when actuators use them.

Repeated scale does not accumulate.

Scale operation on tendon fields: result = default * sample.

Test that history flattening uses term-major ordering. Term-major: [A_t0, A_t1, ..., A_tH-1, B_t0, B_t1, ..., B_tH-1] Time-major: [A_t0, B_t0, ..

record_pre_reset fires before _reset_idx, so action_manager.action is intact. In step(), the sequence is: record_pre_reset(reset_env_ids) <-

Termination terms must return one boolean per environment.

Robot pitched to match a slope → terrain-aware reward ≈ 1.

On flat terrain, terrain-aware and world-up should give same reward.

Robot staying vertical on a slope → terrain-aware reward < 1.

Flat patches flow through generator → importer with correct shape and bounds.