This dataset was created using LeRobot.
Dataset Description
LIBERO extrinsics
Generated from HuggingFaceVLA/libero at commit 86958911c0f959db2bbbdb107eb3e17c5f9c798e. Original state and
action columns are unchanged. The two embedded image columns keep their names
but are stored as AV1 videos at CRF 25.
Pose columns contain [position, rotation, gripper]. Rotations are active,
right-handed R_world_from_end_effector. rot6d stores the first two columns;
rot6d_rows stores the first two rows.
observation.state.ee_rot6d: measured pose with measured gripper position.action.ee_absolute.*: reconstructed OSC command targets with the absolute gripper command, stored in every rotation representation.action.ee_delta.*: the same commands as per-step displacements from the measured pose, in world axes and metric units (p_cmd - p_state,R_cmd @ R_state.T) with the absolute gripper command — LIBERO's native action with the controller scaling (0.05 m, 0.5 rad per unit) undone. Train withdelta_actionsand evaluate withcontrol_mode="relative"plusaction_encoding.observation.state.joint_position: measured Panda arm positions[joint_0, ..., joint_6]in radians, joined exactly from OpenVLA's RLDS demonstrations.observation.intrinsics.static1and.wrist1: 3x3 matrices at every timestamp, stored flat as 9 row-major values[fx, 0, cx, 0, fy, cy, 0, 0, 1].observation.extrinsics.static1and.wrist1: world-from-camera[xyz, quaternion xyzw]at every timestamp, aligned with the stored images.observation.gripper.time_to_closeandobservation.gripper.time_to_open: seconds until the next command transition in the same episode;NaNmeans no remaining event.
meta/stats.json additionally carries per-step statistics for chunk-anchored
actions, under keys
action.ee_absolute.<rotation>.anchored_horizon_50. These
are stats keys, not dataset columns, so normalization ignores them unless a
config asks for them by name. Every statistic is shaped
(50, action_dim); row i describes offset i inside
the chunk, so it broadcasts directly onto a
(batch, chunk_size, action_dim) action tensor.
Row i depends only on the offset, never on the chunk size, so one horizon
covers every shorter chunk: slice rows :chunk_size and assert
chunk_size <= 50. Do not pool the rows — the whole point
is that spread grows with offset. Slicing is not identical to computing stats at
that chunk size directly: only starts with a full 50-step
run inside one episode contribute, so starts in the last
49 frames of an episode never appear at any offset.
That keeps the chunk population identical across rows, and count is that
number of chunks. Every such chunk is counted — these are exact, not sampled.
Anchoring: positions are relative to the chunk's first measured EE position in
world axes, rotations are R_action @ R_anchor.T, and the last column stays the
absolute gripper command. Padded episode tails are never included.
meta/stats.json also carries frame-expressed per-step statistics under
action.ee_absolute.<rotation>.expressed_<frame>_horizon_50
for each of base (native), static1 and wrist1, shaped the
same way. These describe the absolute command targets expressed in a frame's
SE(3) pose frozen at chunk start — positions R.T (p - t), rotations R.T M,
gripper untouched. Absolute poses inherit the frame's own bias, so the frames
differ by far more than scale and a fixed-frame policy should read its own
frame's key: a wrist-expressed chunk fills under a fifth of the pooled range at
offset 0. The frameless key
action.ee_absolute.<rotation>.expressed_horizon_50 pools
all three, one sample per chunk per frame, and stays the right choice for a
labeler that switches frames within an episode. The same chunk population as
the anchored entries contributes; count is chunks times frames when pooled.
Only the wrist rows genuinely vary with the offset — that frame is frozen at chunk start while the arm moves away from it, spreading 9x over 50 offsets. Base and static rows are flat and keep the per-offset shape only so every key loads through the same path.
A third family,
action.ee_absolute.<rotation>.anchored_expressed_horizon_50,
first anchors each chunk on its start state (like the anchored entries) and
then rotates the resulting displacements into the frame's basis — positions
R.T v, relative rotations conjugated R.T M R — pooled over the same three
frames. This is the table for policies that anchor on the live state and
express in a camera frame.
The LeRobot source contains gripper joint positions but not the Panda's seven
arm-joint positions. Those are recovered from openvla/modified_libero_rlds at commit
6ce6aaaaabdbe590b1eef5cd29c0d33f14a08551 by matching every episode's full action trajectory;
the script fails instead of writing joints if any match is not exact.
Published to main and tagged v3.0, which is the revision LeRobot
resolves when a config names no revision. Pin a resolved commit SHA in train and
evaluation configs to keep a run reproducible against later updates.
- Homepage: [More Information Needed]
- Paper: [More Information Needed]
- License: apache-2.0
Dataset Structure
{
"codebase_version": "v3.0",
"fps": 10.0,
"features": {
"observation.state": {
"dtype": "float32",
"shape": [
8
],
"names": [
"state"
],
"fps": 10.0
},
"action": {
"dtype": "float32",
"shape": [
7
],
"names": [
"actions"
],
"fps": 10.0
},
"timestamp": {
"dtype": "float32",
"shape": [
1
],
"names": null
},
"frame_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"episode_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"task_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"observation.images.static1": {
"dtype": "video",
"shape": [
256,
256,
3
],
"names": [
"height",
"width",
"channel"
],
"fps": 10.0,
"info": {
"video.height": 256,
"video.width": 256,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.fps": 10,
"video.channels": 3,
"has_audio": false,
"video.g": 2,
"video.crf": 25,
"video.preset": 12,
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": {},
"is_depth_map": false
}
},
"observation.images.wrist1": {
"dtype": "video",
"shape": [
256,
256,
3
],
"names": [
"height",
"width",
"channel"
],
"fps": 10.0,
"info": {
"video.height": 256,
"video.width": 256,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.fps": 10,
"video.channels": 3,
"has_audio": false,
"video.g": 2,
"video.crf": 25,
"video.preset": 12,
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": {},
"is_depth_map": false
}
},
"observation.state.joint_position": {
"dtype": "float32",
"shape": [
7
],
"names": null
},
"observation.state.ee_rotvec": {
"dtype": "float32",
"shape": [
7
],
"names": null
},
"action.ee_absolute.rotvec": {
"dtype": "float32",
"shape": [
7
],
"names": null
},
"action.ee_delta.rotvec": {
"dtype": "float32",
"shape": [
7
],
"names": null
},
"observation.state.ee_quat_xyzw": {
"dtype": "float32",
"shape": [
8
],
"names": null
},
"action.ee_absolute.quat_xyzw": {
"dtype": "float32",
"shape": [
8
],
"names": null
},
"action.ee_delta.quat_xyzw": {
"dtype": "float32",
"shape": [
8
],
"names": null
},
"observation.state.ee_rot6d": {
"dtype": "float32",
"shape": [
10
],
"names": null
},
"action.ee_absolute.rot6d": {
"dtype": "float32",
"shape": [
10
],
"names": null
},
"action.ee_delta.rot6d": {
"dtype": "float32",
"shape": [
10
],
"names": null
},
"observation.state.ee_rot6d_rows": {
"dtype": "float32",
"shape": [
10
],
"names": null
},
"action.ee_absolute.rot6d_rows": {
"dtype": "float32",
"shape": [
10
],
"names": null
},
"action.ee_delta.rot6d_rows": {
"dtype": "float32",
"shape": [
10
],
"names": null
},
"observation.state.ee_matrix": {
"dtype": "float32",
"shape": [
13
],
"names": null
},
"action.ee_absolute.matrix": {
"dtype": "float32",
"shape": [
13
],
"names": null
},
"action.ee_delta.matrix": {
"dtype": "float32",
"shape": [
13
],
"names": null
},
"observation.state.ee_euler_xyz_fixed": {
"dtype": "float32",
"shape": [
7
],
"names": null
},
"action.ee_absolute.euler_xyz_fixed": {
"dtype": "float32",
"shape": [
7
],
"names": null
},
"action.ee_delta.euler_xyz_fixed": {
"dtype": "float32",
"shape": [
7
],
"names": null
},
"observation.gripper.time_to_close": {
"dtype": "float32",
"shape": [
1
],
"names": null
},
"observation.gripper.time_to_open": {
"dtype": "float32",
"shape": [
1
],
"names": null
},
"observation.intrinsics.static1": {
"dtype": "float32",
"shape": [
9
],
"names": null
},
"observation.intrinsics.wrist1": {
"dtype": "float32",
"shape": [
9
],
"names": null
},
"observation.extrinsics.static1": {
"dtype": "float32",
"shape": [
7
],
"names": null
},
"observation.extrinsics.wrist1": {
"dtype": "float32",
"shape": [
7
],
"names": null
}
},
"total_episodes": 1693,
"total_frames": 273465,
"total_tasks": 40,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
"robot_type": "panda",
"splits": {
"train": "0:1693"
}
}
Citation
BibTeX:
[More Information Needed]
- Downloads last month
- 162