Verify joint pos sensor returns correctly shaped tensor for scalar joint values.

Test keyframe ctrl values match init_state joint positions.

Test ctrl is correctly constructed for an underactuated system.

Test that keyframe joint positions match the configuration.

Test that keyframe joint positions match the configuration.

The XML compiler derives K and R from the nominal triplet.

Test lag=1 returns previous observation.

Test that DC motor produces zero torque at max velocity.

Test that history_length uses consecutive timesteps.

Test that DC motor saturation_effort and velocity limits apply.

Verify input_order='pos_vel' concatenates position then velocity.

Verify input_order='vel_pos' concatenates velocity then position.

Test that multi-joint reshaping works correctly.

Verify network loads from TorchScript file and initializes.

Test reset zeroes history buffers for specified environments.

Test pos_scale, vel_scale, torque_scale are applied correctly.

LightCfg should add lights to spec.

Add offset to light direction, check default + offset.

Set light position to absolute range, check bounds.

Test CSV string conversion utility.

Log-uniform produces values in [lo, hi], skewed toward lower.

Setting the removed ls_parallel option warns instead of erroring.

LuGre friction subtracts a velocity-dependent force after the ``effort_limit`` clamp (mechanical, like cogging).

Managers should not mutate class-based func to instance. When term_cfg.func is a class, _resolve_common_term_cfg instantiates it. Without copying

Values within range, diversity across envs.

ValueError for unknown material name.

Texture repeat randomizes S/T axes for selected materials.

Scale operation uses compiled defaults and both S/T axes by default.

MaterialCfg should add materials to spec.

MaterialCfg.geom_names_expr assigns material to matching geoms.

Test that material colors are properly applied.

Both variants' materials end up in the merged spec, name-prefixed.

Each world's geom_matid points to its variant's material.

A variant slot without a material yields geom_matid == -1 in its worlds.

Regression: merging each visual-compat subgroup must preserve visuals.

Test merging multiple sites combines them correctly.

Test merging a single site returns the site mesh.

Test meshes with texture coordinates (Issue #225).

Step metrics must return one scalar per environment.

Substep metrics must return one scalar per environment.

Test creating a minimal scene with no entities.

When proportions length != num_cols, should fall back to even.

Multiple builtin actuator types can coexist and all use batched path.

Test group with both delayed and non-delayed terms.

Test group with both history and non-history terms concatenated.

BuiltinMotorActuator writes effort targets via batched path.

Physical params pack K = sqrt(Kt*Ke) and R verbatim.

Test MPKPE computes the correct mean global error.

MPKPE must read the global reference, not the drift-cancelled one. Pins issue #1006: setting body_pos_relative_w to match the robot exactly would

Test MPKPE is zero when global positions are identical.

Each frame excludes only its own parent body, not others. body_a has a platform geom directly below site_b. Frame B's rays should skip body_b's o

height_scan works correctly with multi-frame sensors.

Sites at different Z offsets produce proportional distances.

Multi-frame with grid pattern: verifies reshape ordering.

Multi-frame sensor produces correct output shapes.

Sphere (3 geoms) + cone (5 geoms) -> body padded to 5 mesh geoms.

Multi-slot air-time stays per-primary, primaries are exposed by name. Regression for issue #914: previously `num_slots > 1` with `track_air_time`

Verify pattern lists create separate tracking slots for each matched geom. When passing a list of patterns like ["left_foot_geom", "right_foot_geom

Test observation with history_length=3 and flattened.

Test observation with history_length=3 and not flattened.

Test that multiple entities produce a single merged keyframe.

An entity with more than one freejoint is rejected at construction.

Verify multiple sensors can be registered and return correctly shaped data.

Test multiple contact sensors in the same scene.

Test multiple observation terms with different noise configs.

With preserve_order=True, names stay in the user-specified order.

With preserve_order=False, names are reordered to internal order.

Providing full set of names resolves ids and collapses to slice(None).

NaN should be sanitized before entering delay buffer.

NaN should be sanitized before entering history buffer.

NaN checking should work with non-concatenated dict output.

Final-only checking should check concatenated result.

Per-term checking should identify which term has NaN.

Test that nan_detection correctly identifies NaN/Inf environments.

Test that nan_detection is properly logged by termination manager.

Policy='disabled' should pass through NaNs unchanged.

Policy='error' should raise ValueError with context.

NaN guard should capture states and dump when NaN detected.

NaN guard should capture NaN in high-indexed environments beyond max_envs_to_dump.

NaN guard should capture and restore mocap body poses.

NaN guard should create latest symlinks that work correctly.

NaN guard should correctly identify which environments have NaN/Inf.

NaN guard should be disabled by default with no overhead.

NaN guard should only dump once per training run.

NaN guard should only keep last K states in buffer.

NaN guard should save model file alongside state dump.

NaN guard should work with complex robot model.

NaN handling should work correctly with clipping applied.

NaN handling should work independently for multiple groups.

NaN handling should work correctly with scaling applied.

Policy='sanitize' should zero NaNs without warning.

Policy='warn' should log warning and sanitize.

Test that negative infinity is also handled.

3x with fixed value produces same result as 1x (uses defaults).

Test sensor behavior when no contacts occur.

Test that observations work without delay (default behavior).

Test that first call with update_history=True only appends once, not twice.