Base configuration class for actuators.

Configuration for an observation term. The post-processing pipeline mirrors mjlab's ObservationManager exactly: ``compute -> noise -> clip ->

One RoboTwin task — name maps to envs/<name>.py.

Configuration for a reward term. The manager loop computes ``weight * func(env, env_states, **params) * step_dt`` and sums across terms. Per-

Save the observations to images or videos.

Configuration for observation functions.

Base config class for scenes.

Base configuration class for robots.

Configuration for an event term. ``mode`` is one of ``"setup" | "reset" | "pre_step" | "post_step" | "terminate"``. Event functions are dispa

Selector for joints/bodies of a named scene entity.

A registered mjlab task (subset of `Mjlab-*` ids).

World-frame position of a MuJoCo site (works for MJX & raw MuJoCo).

A flat/combined view of multiple controllers.

Configuration for a termination term. ``time_out=True`` routes the term's signal into the time-out (truncation) bucket instead of regular ter

Configuration for reward functions.

Randomize material properties (friction, restitution) for selected bodies.

Configuration for actions in policies. Defines the action type, dimension, and control parameters.

MuJoCo contact data query (works for MJX & raw MuJoCo).

MujocoHandler that does NOT inject a default ground / lights. robosuite's MJCF already supplies its arena floor, lights and textures; MetaSim

Configuration for observations in policies. Defines the observation type, dimension, and format.

RoboTwin's dual-arm embodiment: arx5 arms on an agilex mobile base.

Configuration for a delayed PD actuator.

Configuration for End Effector control (only used if action_space is 'ee_abs' or 'ee_delta').

Cfg for the Franka Emika Panda robot. Args: RobotCfg (_type_): _description_

One robosuite task and how it maps into RoboVerse.

Time-series captured from one rollout.

Wraps RoboVerse environments for RSL-RL OnPolicyRunner compatibility. Works with all RoboVerse environments as they all provide obs_buf and p

Fixed transform profile applied before runtime calibration.

Mimics rsl-rl: actor exposes its trunk under ``actor.mlp``.

Camera intrisics for point cloud creation.

Decoded canonical teleoperation targets for left and right hands.

Optional query to fetch per-body net contact forces for each robot. - For IsaacGym: uses the native net-contact tensor and maps it per-robot in h

Conv1d --> GroupNorm --> Mish

Configuration for a curriculum term. Called from ``_reset_idx`` for the freshly reset env subset. Return value (typically a scalar or dict) i

Configuration for termination functions.

Normalize mean and variance of values based on empirical values.

Track episode returns and lengths without relying on info dict.

Replicates GoogleRobot.check_grasp on a SAPIEN articulation + scene.

One ManiSkill task we want parity with.

A standalone MJCF model shipped by mjlab.

Reproduces GraspSingle...InScene.evaluate success bookkeeping across an episode. Defaults match the coke-can env: require_lifting_obj_for_success

A single rigid collision shape pulled off a native actor.

World-frame rotation matrix of a MuJoCo site (works for MJX & raw MuJoCo).

Small native session step result.

Mirrors mjlab ``ContactSensorCfg``. Only the fields RoboVerse uses are typed; other mjlab kwargs are accepted via ``**`` for cfg-import parit

Normalize mean and variance of values based on empirical values.

Focal Loss, as described in https://arxiv.org/abs/1708.02002. It is essentially an enhancement to cross entropy loss and is useful for classif

Randomize body masses for a given asset.

MLP as used in Vision Transformer (timm.layers)

OSC_POSE (position+orientation, decoupled) for a fixed-base arm + gripper.

ManiSkill ``pd_joint_delta_pos`` controller, generalized over the robot's action layout. ``compute_targets(current_qpos, action)`` returns the ab

Configuration for RoboVerse-SmolVLA integration.

Per-env velocity command sampler config (mjlab UniformVelocityCommandCfg).

Named benchmark camera pose.

Wraps a single ``mjData`` subtree-* field for one body. For ``subtree_angmom``: returns ``(N, 3)`` whole-body angular momentum about the COM

Config for a sensor that wraps one ``mjData`` subtree field for a body.

Callback configuration buckets used by humanoid tasks.

Configuration for callbacks.

images(agentview, eye_in_hand) + proprio -> next CHUNK actions.

Per-step body-contact tracker. RoboVerse port of mjlab ContactSensor. Resolves ``primary_bodies`` to MuJoCo body IDs at first update; then on