Drain action queue. Runs on the main loop thread.

Distribute *num_envs* across buckets proportionally. Every bucket gets at least one when ``num_envs >= len(proportions)``. Remaining slots are al

Quadratic sigmoid tolerance: 1 at x=0, 0 at |x|>=margin.

Material asset name in the merged template's namespace. Source materials are prefixed with the variant name during merge (``<variant>/<material>`

Mesh asset name in the merged template's namespace. Source mesh names are prefixed with the variant name during merge (``<variant>/<mesh>``); aft

Rotation matrix from a (w, x, y, z) quaternion. Independent of MuJoCo/mjlab.

Queue environment-specific debug draw calls into the scene. Acquires ``_sim_lock`` so the clear+requeue is atomic with respect to the backgro

Resolve string-keyed ranges to entity names and recurse.

Compute the union of geom groups across all raycast sensors.

Recompute ``geom_rbound`` and ``geom_aabb`` from current ``geom_size``. Only primitive types (sphere, capsule, ellipsoid, cylinder, box) are handle

Build the 4x4 pseudo-inertia matrix J from MuJoCo body fields. 1. Rotate principal moments into body frame via ``body_iquat``. 2. Apply the paral

Record a frame from the environment. For vectorized environments, only records env[0].

Remove empty ``<default/>`` and ``<default class="..."/>`` elements.

Render non-selected environments into the native viewer scene.

Resample the command for the specified environments.

Resolve a single field's names and IDs. Args: entity: The entity to resolve against. config: Field configuration specifying attribute

Resolve the frame element to its global ID and parent body ID.

Resolve IDs to their corresponding names.

Resolve names to IDs, optimizing to slice(None) when all are selected.

Sample new lags for specified environments. Args: mask: Boolean mask of shape (batch_size,) indicating which envs to sample. Returns:

Track a specific body of a configured asset.

Track the root body of a configured asset.

Setup camera based on config's origin_type.

Configure native viewer camera from viewer config.

Setup appropriate stream for the device.

Determine if CUDA graphs can be used based on device and driver version.

Single-geom cone for simple tests.

Advance exactly one step while paused.

Spherical linear interpolation between two quaternions.

Return a new obs dict containing only the rows at ``ids`` (per group).

Computes the derivative of a sequence of SO3 rotations. Args: rotations: shape (B, 4). dt: time step. Returns: shape (B, 3)

Sphere: 1 visual + 2 collision geoms.

Returns torch.sqrt(torch.max(0, x)) but with a zero subgradient where x is 0. Reference: https://github.com/facebookresearch/pytorch3d/bl

Pin the tracked body's subtree_com to its frame origin (xpos). MuJoCo's tracking camera centers on ``subtree_com[trackbodyid]``, which shifts

Start recording a new video.

Run physics steps for this frame's sim-time budget.

Submit a scene sync job when the update policy allows it.

Temporarily disable the garbage collector. Prevents GC from finalizing stale Warp Graph objects during wp.ScopedCapture, which would record their

Render queued ghosts as one batched handle per (model, body).

Rotate force/torque from contact frame to global frame.

Push camera sensor frames to GUI when needed.

Update the command based on the current state.

Refresh reward/metric plots and histories for the selected environment.

Roll history buffer and insert current contact data at index 0.

Update current lags according to configured policy.

Update the metrics based on the current state.

Validate that provided names and IDs are consistent. Raises: ValueError: If names and IDs don't match.

Validate that all camera sensors share render settings.

Yields ModuleInfo for all modules recursively on path, or, if path is None, all accessible modules. Note: This function is a modified version

Copy history into figure and autoscale y-axis.

Add an arrow visualization using MuJoCo's arrow geometry.

Add tags to the current wandb run. Note: This function stores tags in wandb.config._wandb_tags if the run is not yet initialized, allowing them t

Write processed actions to entity actuator targets. Called on every decimation substep (physics step), not just once per policy step. Each te

Write builtin actuator controls to simulation data.

Delay all command targets with one shared lag. No-op without delay. Every target the policy issues (position, velocity, effort) travels the same

Wraps spec_fn to auto-wrap fixed-base entities in mocap. This enables fixed-base entities to be positioned independently per environment. Returns

Benchmark a single task.

Build the action term from this config.

Build the command term from this config.

Merge a ``VariantEntityCfg``'s variants into a single padded template. Validates that all variants share the same kinematic structure, merges eve

Per-pixel typed segmentation in (B, 2, H, W) format.

Binary mask of the target cube selected by a MultiCubeLiftingCommand. Output shape: (B, 1, H, W) float32.

Capture current simulation state to buffer.

Check for NaN/Inf and dump buffer if detected. Returns: True if NaN/Inf detected and dump occurred, False otherwise.

Clear the wrapped cache to force re-wrapping of arrays. This should be called after operations that modify the underlying warp arrays, such a

Called when the environment closes. Release file handles, flush write buffers, or finalize output here.

Forward to each term's :meth:`RecorderTerm.close`.

Collect variant metadata for entities with mesh variants.