Overlay task / instruction / step / status onto an RGB uint8 frame (returns a new array). Upscales small SimplerEnv frames so the text is legible

Return (policy_obs, privileged_obs). Implemented by subclasses.

Import LIBERO-plus's benchmark module (after fixing the config path).

Triggers ManiSkill envs registration.

A minimal bridge with 14-D vectors + 4 camera RGB streams (tiny 8x8 frames).

A minimal bridge dict shaped like collect_bridge.py --rgb output.

World-frame site position via ``physics.data.site_xpos``.

Return a dictionary with keys "policy" and "critic", each corresponding to a tensor.

Resolve a {regex-or-name: std} spec into a per-joint std tensor. Mirrors mjlab's ``resolve_matching_names_values``. Falls back to a single sc

Resolve a vendored ManiSkill panda URDF, downloading the asset folder from HF if missing. Prefers the local ``roboverse_data/robots/maniskill_pan

Pull the 14-D [L_arm(6), L_grip, R_arm(6), R_grip] state from RoboTwin obs. Prefer the achieved state injected by the eval loop (``robot_join

Run the worker subprocess; return (npz_dict | None, error_str | None).

Check if task is completed. Args: states: Current environment states (TensorState) Returns: Boolean tensor o

Table-overhead camera for the YAM lift_cube scene. The cube scene loads as scene-MJCF (no registered robot), so MetaSim's ``mount_to`` body-m

Pair src->dst Linear layers in order, asserting matching shapes. Returns a list of ``(src_weight_key, dst_weight_key)``. Raises ``ValueError``

Transform approach vectors and in-plane rotation angles to rotation matrices. Input: batch_towards: [torch.FloatTensor, (N,3)]

Convert a RoboTwin bridge pickle into a name-keyed ``*_v2`` dataset.

Mesh RigidObjCfgs for a family's full candidate set (random-object tasks declare them all in the ScenarioCfg; reset activates the per-episode subs

Clear images for next variant.

Close the environment and simulation app.

This long function is unfortunately doing something very simple and is being very defensive: We are separating out all parameters of

Copy a on-disk zarr to in-memory compressed. Recommended

Create an asset converter instance based on target and source types.

Convert URDF files generated by EmbodiedGen into the format required by all simulators. Supported simulators include SAPIEN, Isaac Sim, MuJoCo, I

Return ``{suite_name: [task_name, ...]}`` from LIBERO-plus's benchmark dict. Uses the authoritative ``benchmark.get_benchmark_dict()`` ordering (

Joint velocities for the entity, batched. Shape ``(num_envs, len(joint_ids))``.

Extract state dictionary from handler states (similar to teleop_keyboard). Args: env: Environment with handler scenario: Scenario

Force reset environment to specific state with validation.

Concatenate the three id groups exactly as SimplerEnv's get_obs does for the mask.

For each dt window, choose the first timestamp in the window. Assumes timestamps sorted. One timestamp might be chosen multiple times due to

Return ``(root_quat_xyzw, lin_vel_w, ang_vel_w)`` of the asset's base. Reads from ``env_states.robots[name].root_state`` (RoboVerse layout: p

Get state dict from model, handling DDP wrapper if present.

Get the geometric center of the object in the world frame.

Calculate the geometric center of the object in world coordinates.

Get the geometric center of the object in the world frame.

Compute k nearest neighbors for each query point.

Return {benchmark_name: [env_id, ...]} for all currently-registered passthrough envs.

Logistic barrier function for reward shaping.

Build the native LIBERO ``OffScreenRenderEnv`` for ``(suite, task_id)``. This is the canonical 1:1 construction path (matches LIBERO's README and