Interpolates the motion to the output fps.

Linear interpolation: include the first waypoint, then for each segment add N points (excluding segment start). If waypoints has shape (K, 3) and

A registered object/fixture vs the arena table/floor, by LIBERO's naming. LIBERO's ``On(obj, region-site)`` requires contact with the region's pa

Iterate over valid episode directories that contain both metadata and video.

Map joint indices → qvel/dof offsets (freejoint=6, ball=3, slide/hinge=1).

Map MuJoCo joint indices → qpos offsets (handles freejoint+ball+slide+hinge widths). MuJoCo joints occupy different qpos widths: freejoint=7, bal

Simple k-means clustering algorithm adapted from Karpathy's minGPT library https://github.com/karpathy/minGPT/blob/master/play_image.

Return the installed ``libero/libero`` package dir, or ``None`` if absent. Located via ``importlib`` *without importing* the package, so we can w

List file paths under a remote dataset directory, returning [] on failure.

Return a callable ``policy(actor_obs) -> actions`` from a checkpoint. Handles every checkpoint format used in this repo: * a TorchScript expo

Loads the motion from the csv file.

Find a URDF-object's ``mobility.urdf`` under ``assets/objects/<name>/`` (returns relpath).

Build the pd_joint_delta_pos controller matching the robot's action layout. A mimic gripper adds one extra active DOF beyond the action width (th

Return an mjvCamera that tracks the task's hero body (matches mjlab viewer). Falls back to None (model default free camera) if the task isn't in

Peak displacement of the most-moved task object over the demo (visual-salience score).

Cast `np.float64` action datatype to `np.float32`, keep the others dtype unchanged. See GH#1572 for more information.

Pull the ``assets/<rel>`` path out of a build error (file-not-found / compile).

Rebuild a deterministic actor MLP from an ordered ``mlp.*`` state dict. Matches rsl-rl's ``MLPModel.mlp`` (``Linear, ELU, Linear, ELU, ..., Linea

Format a flat demo action for the native env (a tensor, or a per-agent dict for multi-agent).

Independent native construction (raw OffScreenRenderEnv), no passthrough.

Construct the env exactly like upstream LIBERO-plus eval, independently.

robosuite Lift success, ported: cube center > table top + 0.04 m.

World-frame position of the tracked object.

The ancestor of ``bid`` that is a direct child of worldbody (the object root).

Native LIBERO obs -> (agentview_rgb, eye_in_hand_rgb). The native ``*_image`` obs is already in the same orientation as the demo ``*_rgb`` ar

Get observation using RoboVerse tensor state.

Get observation using RoboVerse tensor state.

LIBERO On(a,b) = b.check_ontop(a). If ``b`` is a region *site* (e.g. ``cabinet_top_side``) LIBERO uses the site's box containment (``SiteObje

Geometric is-A-on-B (ManiSkill xy + z flags), without the static/grasp parts.

Make every saved frame also carry RoboTwin's *achieved* arm qpos. RoboTwin's ``joint_action.vector`` is the PD *command target* (``joint.get_

Strip mocap bodies and align arm motor names with cfg expectations. Finger joints are left alone — the bundle MJCF uses the ``{side}_hand_fin

ManiSkill PegInsertionSide ``has_peg_inserted``: peg head inside the box hole. box_hole_pose = box.pose * [0, 0.0087, 0.0038]; peg_head_pose = pe