Attach to the simulator handler and prepare mass buffers.

Bind handler to the query.

Mjlab ``dr.body_com_offset`` port — perturbs each body's local COM. ``ranges`` is a dict ``{axis: (lo, hi)}`` per the mjlab cfg. Axis 0 = x,

Load the openVLA-OFT model + heads + processor (official path). We can't import ``GenerateConfig`` from openvla-oft's ``run_libero_eval`` bec

Reconstruct rsl_rl ActorCritic MLP from state_dict. Expected keys: ``mlp.{0,2,4,6}.{weight,bias}`` (Linear layers, with ELU between).

Build the scene in-memory and return the compiled MjModel. Skips writing XML to disk so we don't have to chase down meshdir relative-path iss

Build the first native benchmark task migrated from BiDexBench.

Build and serialize the scene to disk; mostly useful for external tools. Note: the saved XML preserves the include + relative meshdir from the

Build the perturbed robosuite env, lazily fetching missing assets; return state.

Build an XML Element for the <robot> tag.

Build an XML Element for the <robot> tag.

Build an XML Element for the <robot> tag, containing <link> and <joint> tags. Returns the root <robot> element (as an xml.etree.ElementTree.E

Calculate optimal camera position based on object bounds.

Calculate the bounding box of all geometries in the model.

GPT must output strictly valid JSON: { "task_name": "...", "task_language_instruction": "...", "robot_involved": [.

GPT merges: - The chosen robot(s) (with known pos, rot, dof_pos from all_robots_data). - The chosen objects (plus any decorative obje

Recursively checks that elements in `values` match the TARGET_SPEC.

Run one task's native-vs-wrapped comparison in a fresh subprocess.

Cleanup for the environment.

Return sanitized 21x3 finger landmarks, or None when unavailable.

Collect ``num_demos`` successful demos by rolling out ``checkpoint``. Args: task / robot / sim: RoboVerse task + robot + backend (as in R

Return ``[B, P]`` bool for primaries that left contact within the last dt seconds.

Compute pair-wise point distances in two matrices. Input: A: [np.ndarray, (N,3), np.float32] point cloud A

Convert RoboVerse demonstrations to LeRobot dataset format.

Convert one ManiSkill episode into a name-keyed multi-agent ``*_v2`` file. Picks the first successful episode when ``demo_index`` is ``None``. Re

Convert state dict to initial state format.

Convert state dict to initial state format.

Convert state dict to initial state format.

Convert state dict to initial state format.

Convert state dict to initial state format.

Convert flat state dict to nested format for handler.set_states.

Convert a single URDF file to USD.

Load to memory.

Copy MJCF file and all its assets to output directory, maintaining relative paths.

Copy texture file to destination directory and return the relative path. Skip PNG textures.

Copy texture file to destination directory and return the relative path.

Copy the three-digit USD scenes into the asset directory.

Create a Msh object from a .msh file.

Create a camera with the specified parameters using USD API directly.

Fill a command with damping-only gains.

Create task environment with lights and camera.

Zero out all fields in the command.

Crops images at the locations specified by @crop_indices. Crops will be taken across all channels. Args: images (torch.Tensor):

Given the latent action, reconstruct the original action. Inputs: latent_action (shape: ... x 1): The latent action to recon

Return the default AVP wrist profile for a simulator backend.

Maintain default position until the A button is pressed.

Open-loop press-down-and-close — works for any gripper task as a probe.