liberoX

LIBERO structural-supervision sidecars aligned to the exact Fast-WAM LeRobot v2.1 release. The four archives retain Fast-WAM's original parquet, metadata, and AV1 videos byte-for-byte and add lossless active-object masks, projected Panda skeletons, camera matrices, and replay provenance.

The labels are training-time privileged targets. They are not intended as policy inputs at deployment.

Contents

Archive	Tasks	Episodes	Frames
libero_spatial_no_noops_lerobot.tar.gz	10	434	53,229
libero_object_no_noops_lerobot.tar.gz	10	457	67,309
libero_goal_no_noops_lerobot.tar.gz	10	433	52,895
libero_10_no_noops_lerobot.tar.gz	10	388	104,280
Total	40	1,712	277,713

LIBERO uses one Franka Panda arm. These skeleton labels are therefore single-arm agent grounding data; bimanual grounding still requires a platform such as RoboTwin.

Fast-WAM compatibility

Extract all four archives into the same directory used by Fast-WAM:

mkdir -p data/libero_mujoco3.3.2
cd data/libero_mujoco3.3.2
for archive in /path/to/liberoX/*_lerobot.tar.gz; do
  tar -xzf "$archive"
done

The resulting names and original trees match Fast-WAM exactly:

libero_goal_no_noops_lerobot/
  data/chunk-000/                   # original Fast-WAM parquet
  videos/chunk-000/                 # original Fast-WAM AV1 videos
  meta/                             # original metadata + annotation metadata
  annotations/chunk-000/
    episode_000000.npz
    episode_000000.json
  validation_report.json
  video_validation_report.json
  fastwam_reference_validation.json
  replay_rgb_validation.json

No Fast-WAM loader change is needed for the original RGB/action path. An annotation-aware loader can join sidecars by episode_index and frame_index.

Annotation schema

Every episode_XXXXXX.npz stores arrays at the original two-camera 512x512 resolution. Camera order is agentview, then robot0_eye_in_hand.

Field	Shape	Meaning
object_mask_packed	[T,2,512,64] uint8	Active-object mask, packed along width
skeleton_packed	[T,2,512,64] uint8	Binary Panda skeleton raster, packed along width
skeleton_points_2d	[T,2,11,2] float32	Projected points in (x,y) order
skeleton_valid	[T,2,11] bool	Positive-depth, in-frame point validity
skeleton_depth	[T,2,11] float32	Camera-space point depth
camera_intrinsic	[T,2,3,3] float32	Camera intrinsics
camera_extrinsic	[T,2,4,4] float32	World-to-camera transform
camera_to_world	[T,2,4,4] float32	Camera-to-world transform
source_action_index	[T] int32	Index in the official source demonstration
gripper_state	[T,2] float32	Fast-WAM-aligned gripper qpos

Binary arrays use numpy.packbits(..., axis=-1, bitorder="little"):

import numpy as np

with np.load("annotations/chunk-000/episode_000000.npz") as episode:
    object_mask = np.unpackbits(
        episode["object_mask_packed"],
        axis=-1,
        count=512,
        bitorder="little",
    ).astype(bool)

For Fast-WAM's LIBERO layout, resize each binary camera frame to 224x224 with nearest-neighbor interpolation and concatenate horizontally:

agentview 512x512 -> 224x224
wrist     512x512 -> 224x224
horizontal concat -> 224x448

Do not use bilinear interpolation for binary labels.

Label definitions

Active-object mask. The mask is the union of directly manipulated task instances. Roles are inferred from LIBERO BDDL goal predicates: the first predicate argument is active, while later arguments are targets. Region names resolve to their owning instance. Masks are simulator instance masks, not SAM pseudo-labels.

Panda skeleton. The 11-point tree contains robot0_link0 through robot0_link7, robot0_right_hand, and both gripper fingertip bodies. The raster is an amodal projected kinematic tree on a black canvas.

Before every label render, Panda arm and gripper qpos are set from the matching Fast-WAM parquet row. The simulator replay qpos are restored before the next source action. This keeps wrist-camera geometry and skeletons tied to the released Fast-WAM observation rather than accumulated contact-replay drift.

Replay protocol

• MuJoCo 3.3.2, robosuite 1.4.0, LIBERO commit 8f1084e;
• environment seed 0;
• restore official LIBERO HDF5 states[0];
• execute 10 settling actions [0,0,0,0,0,0,-1];
• apply OpenVLA's 1e-4 no-op filter without executing removed actions;
• capture labels before each retained action;
• preserve the released Fast-WAM episode and frame ordering.

meta/replay_manifest.jsonl records the unique official HDF5 demo match, retained source-action indices, and action SHA256 for every episode.

Validation

Each extracted dataset includes machine-readable reports covering:

• LeRobot episode, parquet, video, manifest, and annotation counts;
• NPZ shapes, dtypes, frame alignment, packbits round trips, and camera inverse checks;
• AV1 codec, 512x512 dimensions, 20 FPS, and per-episode frame counts;
• byte size and SHA256 of every original Fast-WAM archive member;
• full replay-to-video RGB audits for both cameras.

Top-level checksums.sha256 contains SHA256 digests for the published files. Generation and validation source is provided in liberoX_generation_tools.tar.gz.

Sources

• Fast-WAM
• Fast-WAM paper
• LIBERO
• OpenVLA LIBERO regeneration script