Curriculum reward that gates lifting bonus on reaching progress. Returns reaching * (1 + bringing), where both terms are Gaussian kernels over po

Config with a stateless class-based reward.

Environment step size (physics_dt * decimation).

Convert a string representation of a slice to a slice object. Args: s: The string representation of the slice. Returns: The

Access the underlying wrapped struct.

Create a network that subtracts second half from first half of inputs. For input [a, b, c, d], returns (a + b) - (c + d). This is sensitive to in

Copy env state to viewer; update reward figures; render other envs.

Synchronize environment state to viewer.

Sync mouse perturbation to sim via ``qfrc_applied``. Mouse perturbation forces are converted from Cartesian body space (``xfrc_applied``) to

Synchronize viewer state to environment (e.g., perturbations).

Geom IDs of the target cube per env. Shape: (B, K).

Local indices of targets controlled by this actuator.

Target IDs for the controlled transmission.

Names of targets controlled by this actuator.

Target names for the controlled transmission.

Target position in EE frame.

Randomize tendon armature. Triggers ``set_const_0``.

Randomize tendon damping.

Randomize tendon friction loss (tendon_frictionloss).

Tendon lengths. Shape (num_envs, num_tendons).

Randomize tendon stiffness.

Tendon velocities. Shape (num_envs, num_tendons).

Register a terrain preset into ALL_TERRAIN_PRESETS.

Values set directly within range, diversity across envs.

Verify accelerometer reads non-zero acceleration when robot is on floor.

VideoRecorder accepts a string path for video_folder.

Fractional budget carries across ticks.

A pos error big enough to make kp*err exceed effort_limit must be clamped at the joint, not allowed to ride through the unbounded element.

Test that reset clears action history for all or specific environments.

Test that action, prev_action, and prev_prev_action track history correctly.

Each action type writes to correct entity.data field.

We deliberately leave per-element forcerange unset; the limit lives on the joint. Inspection of actuator_force[i] thus shows the unclamped value.

Error message includes namespace hint when name exists in another namespace.

Values = default + sample, within expected bounds.

dump_yaml must be called before runner creation. The runner mutates agent_cfg in-place (e.g. resolve_symmetry_config injects non-serializable obj

Verify contact sensors track time spent in/out of contact when enabled. Tests the track_air_time feature which monitors how long each contact point

All builtin actuator types preserve XML armature/frictionloss.

Each preset can generate terrain without error.

All tasks should be loadable in play mode.

All registered tasks should be loadable without errors.

A group whose every term is None is skipped rather than raising.

Largest-remainder rounding must always allocate exactly nworld worlds.

A zero-weight variant gets zero worlds; the rest split nworld.

Appending wrong batch size raises error.

Impulse is applied and cleared after duration expires.

The first call after init/reset enters cooldown, not an immediate impulse. Regression test for #973.

reset() zeros forces and resets internal state.

body_point_offset adds cross-product contribution to torque.

Zero rotation delta should return finite values and preserve input pose.

Joint armature must come from the position element only; double-applying would silently double dof_armature.

Articulated variants with matching topology pass validation (build still rejects via floating-base check; this verifies validation itself does not

When assignment_fn is set, it dictates the per-world variant indices.

Per-world independent RNG draws make world W's variant a function of W alone, independent of nworld.

Test that metadata can be attached to ONNX models.

After auto_reset=False, calling reset(env_ids=...) resets those envs.

Terminal obs (auto_reset=False) differs from post-reset obs (auto_reset=True).

With auto_reset=False, done envs are NOT reset and obs is the terminal state.

Raw env should reject another step until done envs are explicitly reset.

Example: run your own training loop against an auto_reset=False env. The pattern is: 1. After step(), derive done_ids from terminated | truncat

With auto_reset=True (default), done envs are reset during step.

Only the specified axes are modified; others stay at defaults.

Same V, joint moving at omega: tau = K * (V - K * omega) / R.

After resetting a row, the next append back-fills its entire history for that row.

First-push backfill works for >2D data whose trailing dims differ from batch. Regression for #1045: the backfill broadcast hard-coded a 3D conditio

BaseAction: processed = raw * scale + offset.

BaseAction: clip clamps only matched actuators; others stay unclipped.

BaseAction.reset() zeros raw_action for specified env_ids only.

Base RecorderTerm hooks do nothing; __call__ raises NotImplementedError.

Verify that contact sensors detect collisions between a falling box and ground. Tests that when a box is placed just above ground and simulation st

Verify rays detect the ground plane and return correct distances.

Multiple envs with different orientations get independent rewards.