Envs with different biases should have identical physics when compensated. If env0 has bias=0 and env1 has bias=0.3, and we want both to reach phys

body_ipos modification triggers correct recomputation.

Modify body_ipos: GPU recompute matches CPU mj_setConst.

dof_invweight0 changes after body_mass modification.

Modify body_mass: GPU recompute matches CPU mj_setConst.

body_subtreemass is correctly recomputed after body_mass change.

Test contact detection with body mode.

Yaw-only perturbation: roll/pitch of result match default.

WarpBridge caches wrapper objects.

Slice assignment preserves memory addresses (CUDA graph safe).

Direct assignment raises AttributeError with helpful message.

WarpBridge wraps warp arrays as TorchArray.

Builtin actuator resolves joints through internal attach prefix.

Builtin actuators use batched path, custom actuators use compute().

Controls are written correctly when actuator order differs from joint order.

BuiltinPositionActuator writes position targets to ctrl.

Verify BuiltinSensorCfg throws error when name conflicts with XML sensor.

Verify that reset() invalidates the cache.

Verify that update() invalidates the cache.

Test that observation cache is invalidated when environments are reset.

Set fovy to absolute range, check bounds.

Per-env cam_intrinsic DR produces different images across envs.

Add offset to camera positions, check default + offset.

Zero RPY range preserves default quaternion.

CameraCfg should add cameras to spec.

clone_data=True returns independent tensors.

Camera sensor can create a new camera (not wrap existing).

Camera sensor can create a new camera attached to a body.

Depth data has correct shape [num_envs, height, width].

Subset of envs randomized, others unchanged.

Both RGB and depth can be requested simultaneously.

All three data types can be requested simultaneously.

RGB data should contain non-zero values when scene has objects.

RGB data has correct shape [num_envs, height, width, 3].

Segmentation contains typed MuJoCo ids for geom hits and background.

Segmentation data has correct shape [num_envs, height, width, 2].

Scene creates SensorContext when camera sensors are present.

Wrapping an existing camera can override its fovy.

Capped resets to False each tick; only True when deadline hit.

Deadline hit with no remaining budget is NOT capped (transient stall).

Basic append and chronological retrieval (oldest -> newest).

First append back-fills whole history for each batch row.

Overwrites oldest once capacity reached.

Reset clears values and counters for specified batch rows.

Class-based term's reset() is called with the correct env_ids.

Test that class-based reward terms are tracked and have reset called.

as_onnx() with SpatialSoftmaxCNNModel matches the original model.

SpatialSoftmaxCNNModel exports to a valid ONNX file.

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

Cogging is added *after* the forcerange clamp (MuJoCo's intentional model: ``effort_limit`` bounds electromagnetic torque, cogging is mechanical).

At ctrl=0 (no electromagnetic torque), qfrc_actuator equals the cogging term ``A * sin(Np * q + phi)`` evaluated at the joint angle.

``cogging=(A, Np, phi)`` packs into ``biasprm[0:3]``.

Collinear points don't define a plane, should fall back to [0, 0, 1].

CollisionCfg should set basic collision properties.

CollisionCfg should support dict-based field resolution.

CollisionCfg should disable non-matching geoms when requested.

Go1 should have 4 foot collision geoms.

CollisionCfg should set margin, gap, and solmix on matched geoms.

CollisionCfg should support dict-based margin, gap, solmix overrides.

CollisionCfg should support regex pattern matching.

CollisionCfg should validate parameters.

Test compiling an empty scene.

Test compiling a scene with entities.

Verify _compute_data is only called once when cache is valid.

Test that compute(update_history=False) returns cached result. This test verifies the fix for the double-push bug where calling compute() multipl

Test observation with constant delay (min_lag = max_lag = 2).

Verify all contact sensor output fields have correct shapes and values. Tests that force, torque, dist, pos, and normal fields are properly populat

env_switch requests scene update only when contacts are enabled.

NullRecorderManager has no terms: `in` is always False and get_term raises.

`in` reflects whether a term name is registered on the manager.

Total terrain_types length must always equal num_envs.

Test box site mesh creation.

Test capsule site mesh creation.

Test cylinder site mesh creation.

Test ellipsoid site mesh creation.

Test that all-zero RGBA falls back to gray.

Test sphere site mesh creation.

Critic observations should always have corruption disabled.

Test that entity.indexing.ctrl_ids are in actuator definition order. ctrl_ids follow actuator definition order for simplicity. ONNX export builds

Demonstrate that ctrllimited=True clips control internally (the bug). This test shows what would happen with the bug: commanding at vs beyond the

current_length counts per-batch valid frames and clamps to max_len.

A user-defined Distribution works end-to-end.

Test that custom Entity subclasses work through the scene.

A user-defined Operation works end-to-end.

Verify cutoff parameter propagates to MuJoCo model.

Verify sensor data reflects physics state after step + update cycle.

Each world's geom_dataid points to its variant's meshes.

DC motor transitions correctly at corner velocity where curves intersect.

Continuous effort_limit constrains output below saturation_effort.

DC motor torque varies linearly between zero and max velocity.

DC motor handles negative velocities symmetrically.

DC motor produces full saturation_effort at zero velocity.

DcMotorActuatorCfg warns when effort_limit > saturation_effort.

DcMotorActuatorCfg warns when effort_limit is inf.

DC motor produces zero torque at maximum velocity.