qfrc_external should match J^T * xfrc_applied computed by CPU MuJoCo.

Fixed yaw=pi/4 matches manual quat_mul(q_yaw, q_default).

All output quaternions must have unit norm (within 1e-6).

All ranges (0, 0): result equals default quaternion.

R-MPKPE reads the re-anchored reference, not the global one. Setting body_pos_w to match the robot exactly would yield zero error if it were (inc

R-MPKPE is zero when re-anchored positions are identical.

randomize_encoder_bias should sample values within specified range.

x/y range filters restrict where patches can land.

Verify world alignment keeps rays fixed.

Verify yaw alignment ignores pitch/roll.

Test yaw alignment handles 90 degree pitch singularity correctly. With yaw alignment, rays should maintain their pattern regardless of body pitch

Verify rays rotate with the body frame.

Verify rays work with non-default direction.

Verify ValueError is raised for invalid frame type.

Verify parent body is excluded from ray intersection when configured.

Verify rays can be attached to a geom frame.

Verify rays can be attached to a site frame.

Verify grid pattern generates correct number of rays.

Verify hit_pos_w returns correct world-space hit positions.

Verify include_geom_groups filters which geoms are hit.

Verify hits beyond max_distance are reported as misses.

Verify rays that miss return distance of -1.

Verify surface normals point upward when hitting a horizontal floor.

Test read properties are stale until forward() is called. Demonstrates why event order matters and why forward() is needed between writes and rea

IntEnum max() gives strongest level.

Non-DR events don't trigger recompute.

record_post_reset is called after obs are computed, so obs_buf already holds the new-episode initial observation for the reset environments.

record_pre_reset fires before record_post_step within a single step.

obs_buf at record_pre_reset is the previous step's observation, not the post-action terminal state. It is the input obs the agent used to pick the t

Variants with different grandchild counts fail recursive validation.

reduce='last' reports the last step's value, not the episode average.

All three reduce modes work correctly in the same manager.

reduce='max' reports the highest value seen during the episode.

After reset, max tracking restarts from -inf.

reduce='max' with per_substep tracks the max of step-averaged values.

Mixed reduce modes in the same manager report correctly.

Verify reduction modes correctly aggregate multiple simultaneous contacts. Tests "none" (no filtering), "mindist" (closest contact), "maxforce" (st

Test reduction modes with multiple simultaneous contacts.

Verify regex patterns correctly match multiple geoms with similar names. Tests that a pattern like ".*foot_geom$" matches all geoms ending with "

Encoder bias must not affect the target: target = current_pos + delta.

With nonzero action, targets shift by action * scale from current.

Setting offset on RelativeJointPositionActionCfg raises ValueError.

With zero action, targets equal current_pos.

reset(None) zeros the entire buffer and counters without indexing.

reset_environment() calls policy.reset() if available.

Test that reset without env_ids clears all environments.

Reset clears delay buffers so old values don't leak through.

Test that reset clears delay buffer and restarts lag constraint.

reset_joints_by_offset applies offsets and respects joint limits.

reset_joints_by_offset must not clamp unlimited joints to [0, 0].

Test that reset with specific env_ids only resets those envs.

Test that reset envs get backfilled on next update.

Test partial reset actually resets only specified environments.

reset_environment() works fine when policy has no reset method.

Unknown distribution string raises ValueError.

Unknown operation string raises ValueError.

Test that RewardManager converts NaN/Inf reward values to zero.

Test that RewardManager converts negative infinity to zero.

Test that scale_by_dt defaults to True for backward compatibility.

Test that rewards are not scaled by dt when scale_by_dt=False.

Test that rewards are scaled by dt when scale_by_dt=True (default).

Reward terms must return one scalar per environment.

Concentric rings: 1 center + 4 + 6 + 8 = 19 rays.

Ring pattern hits floor and ring offsets lie on a circle.

Ring with include_center=False omits the center ray.

Ring with 8 samples + center = 9 rays.

Tests that all robots in the asset zoo compile without errors.

Test root state can be written and read from simulation.

Test angular velocity converts from world to body frame internally. The API accepts angular velocity in world frame, but MuJoCo's qvel stores it

Test reading and writing root velocity is a no-op (world frame). Verifies that the API is consistent: if you write a velocity, read it back, and

Export terrain with hfield via Entity.write_xml, reload as MjModel.

Rough velocity play tasks should have terrain curriculum disabled.

Rough velocity tasks should have generator terrain.

Rough velocity training tasks should have terrain curriculum enabled.

Old checkpoints without env_state should load without crashing.

MjlabOnPolicyRunner should save and restore common_step_counter.

Clearing sameframe shortcuts aligns host mj_forward with variant.

Scalar material fields randomize per selected material.