Test that observations work without history (default behavior).

No stone center should fall inside the platform.

When no per_substep terms exist, compute_substep is a no-op.

Test noise with 'abs' operation (replaces data with bias).

Test noise with 'add' operation.

Test that noise is applied when enable_corruption=True.

Test that noise is NOT applied when enable_corruption=False.

Test noise with 'scale' operation.

Test that tensor conversion is cached across multiple calls.

Test noise with per-dimension bias (tuple).

Delayed non-builtin actuators remain in custom_actuators.

None terms in config are skipped without error.

Every ``from mjlab... import X`` in a notebook must resolve.

NullMetricsManager doesn't crash and returns empty dict on reset.

Test behavior with num_slots > 1.

The observation fn reflects IMU site tilt; the entity-data version does not. With the base upright but the IMU site rolled 30 deg about x, projecte

Terms shared between policy and critic groups get independent settings.

as_onnx() model produces the same output as the full actor with normalization.

as_onnx() works when normalization is disabled.

_OnnxMotionModel clamps time_step beyond motion length to last frame.

_OnnxMotionModel actions output matches calling the actor directly.

_OnnxMotionModel returns the motion data at the requested time step.

Exported .onnx file produces the same outputs as PyTorch via onnxruntime.

Passing an Operation instance works identically to the string.

Shorter variant's padding geom slots have dataid == -1.

Values set directly within range, diverse across envs.

Values = default + sample; no accumulation after 3 calls.

Only the targeted axis changes; others stay at their original values.

Targeting pair by name only modifies that pair.

ValueError for unknown pair name.

isotropic=True with roll axes mirrors roll2 = roll1; tangent axes unchanged.

isotropic=True with dict-int ranges containing roll axis triggers roll2 = roll1.

isotropic=True + shared_random=True: all pairs same value, tangent2 == tangent1.

isotropic=True mirrors tangent2 = tangent1; spin and roll axes unchanged.

axes=[0, 1] randomizes both tangent components; axes 2-4 stay unchanged.

Randomizing a subset of envs leaves the others unchanged.

Values = default * sample; no accumulation after 3 calls.

All pairs in the same env get the same value; envs differ.

Dict with pair name patterns randomizes the matched pairs correctly.

Randomizing subset of envs leaves others unchanged.

Half invalid points with junk z values → still correct normal.

Setting only armature preserves XML frictionloss, and vice versa.

Partial reset returns zeros until next append backfills.

Patches must not land near a sharp height discontinuity.

No patch center should be within patch_radius of the heightfield edge.

PD gains scale/abs on Builtin, Xml, and Ideal actuators.

dr.scale / dr.abs Operation objects produce the same result as strings.

Independent values per environment.

Operations other than scale/abs raise ValueError.

Both targets nonzero: qfrc = kp * pos_target + kd * vel_target.

Substep terms are averaged across substeps; step terms evaluated once.

Test conversion performance with a complex model.

Verify that Perlin noise terrain generates a valid TerrainOutput.

Verify that higher resolution increases pixel count.

Verify Perlin noise terrain with a border.

Verify pinhole pattern generates width * height rays.

Verify from_intrinsic_matrix creates correct config.

Verify pinhole pattern can be created from MuJoCo camera.

Verify pinhole pattern works with MuJoCo camera using fovy (not sensorsize).

Play mode tasks should disable push_robot event.

Play mode tasks should have infinite episode length.

Play mode tasks should have observation corruption disabled for policy.

Play and training configs should have matching action structure.

Play and training configs should have matching observation structure.

JointPositionAction should subtract encoder_bias from commands.

Verify that position actuators allow commanding setpoints beyond joint limits. This is the core fix: position actuators have ctrllimited=False, all

BuiltinPositionActuator writes position targets via batched path.

Commanding to ctrlrange limits produces forces matching forcerange.

Verify that actuator forces are clipped to effort_limit, not control limits. While control setpoints aren't clipped, forces must still be limited t

kd=0, no Vmax clamp: tau = K * kp * (target - q) / R.

Huge position error -> PID voltage saturates at Vmax.

Zero vel target: qfrc = kp * pos_target.

Posture weight biases joints toward target when task error is zero.

Every terrain preset must generate compilable MuJoCo geometry across the full difficulty range. Difficulty 0 is exercised explicitly because curricu

With large num_envs, actual distribution should track proportions.

alpha-only: mass and inertia scale by e^{2α}, ipos unchanged.

d_range=(r,r) gives same result as d1=d2=d3=(r,r).

3x with alpha=0.1 gives same result as 1x (uses defaults each time).

Randomizing only env 0 leaves env 1 unchanged.

After wide perturbation, J ≻ 0, mass > 0, triangle inequality holds.

t1-only: mass unchanged, ipos[0] shifts by exactly t1.

Zero perturbation preserves physics for a body with non-trivial ipos/iquat. Since ``eigh`` returns eigenvalues in ascending order while MuJoCo may

All ranges (0, 0): mass, ipos, inertia, iquat equal defaults (1e-5).

At difficulty 0 the step height collapses to 0, but the flat border frame must still be generated as solid, non-degenerate geometry (regression for

qfrc_actuator should equal actuator_force projected through joint gear.

qfrc_actuator should expose only articulated joint DoFs.