Resolve mjlab-style ``{name-or-regex: value}`` to a per-name list.

Replay a demo episode through the shipped task from the demo initial state (1 or N robots). Returns ``(states, success, track_positions)`` where

Place the oracle's sim at the handler's current qpos/qvel and forward.

Place the handler's physics at the oracle's current qpos/qvel.

Validate that current state matches expected initial state for critical objects.

Validate that current state matches expected initial state for critical objects.

Validate that current state matches expected initial state for critical objects.

Return [(label, task_name)] for clean + dynamics-preserving perturbations. "clean" = the neutral-view / neutral-init variant (no actual camera/in

A handler.physics-like view over a robosuite MjSim (raw structs).

Given a <geometry> element and a dictionary describing shape, add the appropriate child (<box>, <sphere>, <mesh>, etc.).

Given a <geometry> element and a dictionary describing shape, add the appropriate child (<box>, <sphere>, <mesh>, etc.).

Given a <geometry> element and a dictionary describing shape, add the appropriate child (<box>, <sphere>, <mesh>, etc.).

Apply one stochastic augmentation step (HDRI + optional materials). Camera is intentionally untouched — replays must be pixel-aligned.

Apply all randomizers with global deferred visual flush. New Strategy (Performance Optimized): - Set global defer flag on handler to block AL

base_env._load_agent: set_articulation_render_material(robot, specular=0.9, roughness=0.3).

Set the gripper finger friction material (WidowXDefaultConfig.urdf_config).

Body-frame linear velocity of the asset's base. Shape ``(num_envs, 3)``.

Bind handler to the query.

Build a format-faithful multi-agent dataset: ``{robot_name: [demo, ...]}``. Args: num_steps: Number of trajectory steps per agent. R

ManiSkill ``gym_utils.clip_and_scale_action`` (float32, matches the torch path).

Clip action to [-1, 1] and scale according to a range [low, high].

Close the underlying envs.

Close the underlying MetaSim handler.

Shut down the pygame subsystem.

Mjlab parity: ``[B, P]`` bool — primaries that landed within the last dt seconds.

Compute the training loss using the flow matcher. Args: model: The flow network (e.g., ConditionalUnet1D or FlowTransfor

Set MUJOCO_GL=egl (or osmesa fallback) for headless rendering.

Convert a teleop control command into MetaSim handler actions.

Given state and delta action sequence generate equivalent goal position and orientation for each step keep the original gripp

Create input and output transforms for RoboVerse-SmolVLA integration. Args: config: Configuration object. If None, uses default conf

Detect collision of grasps. Input: grasp_group: [GraspGroup, M grasps] the grasps to check

Compare two trajectories. Assumes both rollouts share the same T, nq, nv, control sequence.

Render instructions and joint selection status.

dummy_drawer procedural white floor as a visual-only static box (handler-loaded).

Ensure environment is in clean initial state with intelligent validation.