env_switch respects enabled flag; queue clears + calls update_visualizers.

Regression test for #850: deepcopy should not double-prefix sensor name.

Creating a manager must not mutate the caller's cfg dict. _resolve_common_term_cfg replaces cfg.func with the instantiated object in-place on the

With nonzero fixed lag, a new target takes lag steps to appear.

Constant lag returns observation from N steps ago.

Test that delay buffer is not pushed twice per control step. This tests the specific bug where calling compute() twice per control step (e.g., in

Compute before append raises error.

Per-env mode allows different lags per environment.

Reset clears all environments.

Reset with batch_ids only resets specified environments.

Shared mode uses same lag across all environments.

Update period controls how often lags are resampled.

Input validation catches invalid parameters.

Zero lag returns current observation.

A 2-step lag should make position-mode torque reference step-0 target.

Test that delay preserves observation dimensions.

The shared delay buffer must lag pos and vel together.

Test that delay is applied before history.

Test that scaling is applied before delay.

Test that reset clears the delay buffer.

Test that set_lags sets lag values on all delay buffers.

Test that setting lags actually changes the delay behavior.

Test that set_lags clamps values to the configured lag range.

Test that set_lags gets overwritten when delay_hold_prob < 1.0.

Test that set_lags can set lag values for a subset of environments.

Test that delayed builtin actuator delays position targets.

Fused delayed builtins write controls to sim correctly.

Test that delayed ideal actuator delays position targets.

Feedforward effort targets share the same delay as position targets.

Velocity targets share the same delay as position targets. Regression test: the velocity reference used to bypass the delay buffer, so the dampin

Per-world body_mass matches independently compiled variant models.

{0: (lo, hi), 1: (lo2, hi2)} per-axis ranges.

Delayed builtins with different delay configs get separate groups.

Passing a Distribution instance works identically to the string.

dof_armature modification updates dof_invweight0.

Modify dof_armature: GPU recompute matches fresh MjSpec.compile().

``dr.body_mass`` scale must use each variant's own baseline. This is the load-bearing claim of ``_per_world_default_fields``: scaling body_mass o

``scale`` multiplies the configured ``effort_limit`` (50.0) by the sample.

@requires_model_fields functions populate manager.domain_randomization_fields.

Buffer preserves dtype and device.

Envs with different episode lengths get correct per-step averages.

Unknown end-effector names raise instead of silently scoring zero.

Test EE orientation error detects rotations.

Test EE position error only uses specified bodies.

forcerange clamps the algebraic torque output.

effort_limit=None: jnt_actfrclimited stays 0 on the targeted joints.

Effort limits scale/abs on Builtin, Xml, and Ideal actuators.

dr.abs Operation object produces the same result as the string.

Independent values per environment.

Operations other than scale/abs raise ValueError.

Scale operation uses defaults (no accumulation).

Test electrical_power_cost on robots with passive joints. This test verifies that: 1. The reward correctly identifies and uses only actuated join

Active groups coexist with empty ones; only the empty group is dropped.

A group declared with no terms is skipped rather than raising.

encoder_bias can be set differently per environment.

encoder_bias should be zeros with shape (num_envs, num_joints).

Test that all EntityData properties can be accessed without errors.

Test that EntityData.clear_state() zeros out all target buffers.

Test that EntityData.clear_state() can reset specific environments.

Test accessing entities through dictionary.

Test accessing entities through __getitem__.

Test entity type properties and element counts.

Entity XML <option> flags are not propagated by MjSpec.attach().

Each env_origin must correspond to a valid (level, type) in origins.

Test env_origins property without terrain.

Build a full ManagerBasedRlEnv with variants; step without crashing.

Compute for N steps, reset a subset, verify averages and zeroing.

Verify ValueError is raised when ref provided for unsupported sensor type.

Verify ValueError is raised when obj is not provided for required sensor type.

Verify ValueError is raised when wrong obj type is used for body sensor.

Verify ValueError is raised when wrong obj type is used for joint sensor.

Verify ValueError is raised when wrong obj type is used for site sensor.

Error paths: wrong batch, pre-append access, and bad key size.

When num_envs >= num_cols, every column gets at least 1 robot.

Verify exact name exclusion removes specific geoms from contact detection. Tests the ergonomic feature where passing exact geom names like ("leg1

Verify exact names and regex patterns can be mixed in exclude lists. Tests that exclude tuples can contain both exact names ("trunk_collision") a

Verify regex exclusion patterns filter out groups of similarly-named geoms. Tests that patterns like ".*thigh_collision\\d+" can exclude all thigh

Expanding cam_intrinsic disables precomputed rays on the render context.

Verify CUDA graph is recreated after expand_model_fields.

Explicit values override XML armature/frictionloss.

Explicit 0.0 overrides XML values (distinguishes None from 0.0).

runner.export_policy_to_onnx() produces a valid ONNX file.

Test forces translate, torques rotate, and forces can be cleared.

Test external forces can be cleared.

Test applying force to specific body in articulated system.