Simple gripper control: close when near object.

Return x as an HxWx3 uint8 array, or None if it isn't an image.

Resolve the un-perturbed base BDDL file by stripping all perturbation tokens. The LIBERO-plus benchmark task list contains only *perturbed* entri

Parse ``cfg.events`` (optional) into per-mode callable dicts. Format mirrors beyondmimic's callback dicts: each entry is either a cal

Reconstruct the rsl_rl actor MLP from an ``actor_state_dict``. rsl_rl saves the actor as ``mlp.{0,2,4}.{weight,bias}`` (Linear layers with EL

Map each joint to its exact mjlab motor spec.

Map every trailing path-suffix of each local asset file to its abspath(s). LIBERO object meshes/textures use NON-unique generic filenames — every

Build a RoboVerse task env with a single RGB(+depth) camera for rendering. Reuses the same task-discovery + scenario path as ``roboverse_lear

Construct the vendored native env for a task, resolving assets from roboverse_data.

Map env_id -> body ids aligned with self.body_names for Newton.

Build per-env shape index lists for Newton.

Ensure the MJCF offscreen framebuffer is at least ``width`` x ``height``. LIBERO's stored MJCF caps the offscreen buffer at 480 (no explicit

Calculate raw reward components.

Record every scene object's initial world pose (rigid actors + articulations).

Return (state_obs|Δ|, image_obs|Δ|, reward|Δ|, done_match) across the traj. ``state_obs`` = every non-image obs key (proprio / object poses / etc

Complete trajectory tracking initialization (after super().__init__()).

Complete trajectory tracking initialization (after super().__init__()).

Compute effort from actions using PD control.

Computes the frame blend for the motion.

Compute if object is grasped (requires 5 stable frames based on distance only).

Compute if object is grasped (requires 5 stable frames based on distance only).

Compute if object is grasped (requires 5 stable frames based on distance only).

Compute all observations of a given group and concatenate them into a single tensor.

Computes the velocities of the motion.

Configure the action and observation spaces for the Gym environment.

Normalize live and offline renderer options before launching physics.

Contact between any body in ``root``'s subtree and body ``other`` (the object).

Write one demo_<id>/ dir from a bridge pickle; return the frame count (0 if skipped). Writes ``rgb.mp4`` for the primary ``camera`` (back-compat)

Pick a deterministic ctrl sequence that exercises the model. - cartpole: small sinusoid on the single motor. - everything else: zero ctrl (pa

Yaw angle of all envs (from registered robot root quat).

Prefer the live RoboTwin collection dir; fall back to the vendored slim trajectories so replay works after ~/projects/robotwin is deleted.

Front 3/4 PinholeCamera matching collect_demo's default (rgb+depth).

Return list of per-step diffs between two policy trajectories.

Load via dm_control, export, return the path to the exported root XML. Caller is responsible for keeping the parent tempdir alive (we keep the

Fetch the SimplerEnv asset subtrees from HF into the default roboverse_data dir. Returns the populated root, or ``None`` if the download is unava

Emit visual-only primitives (no physx body) as fixed-pose objects. ``create_visual_box`` (e.g. stamp_seal's stamp-target marker) builds a render-

Point ``LIBERO_CONFIG_PATH`` at a config that resolves to LIBERO-plus. Returns ``True`` if a usable config is in place. Honors a pre-existing