|
Download README.md from PaddyDataCollection/cable_sorting_sample_full: direct link, hf CLI and curl.
- Browser
- Download file 47 kB
-
https://huggingface.co/datasets/PaddyDataCollection/cable_sorting_sample_full/resolve/main/README.md
- Command line
-
hf download hf://datasets/PaddyDataCollection/cable_sorting_sample_full/README.md
-
curl -L -o README.md https://huggingface.co/datasets/PaddyDataCollection/cable_sorting_sample_full/resolve/main/README.md
47 kB
---
license: cc-by-4.0
pretty_name: "Cable sorting sample, full: every recorded stream, lossless (bimanual xArm7)"
task_categories:
- robotics
tags:
- LeRobot
- robotics
- teleoperation
- bimanual
- manipulation
- xarm
- depth
- stereo
- lossless
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files: data/*/*.parquet
---
# Cable sorting sample, full (lossless)
4 teleoperated episodes of one bimanual task, *Pick cable from one bin and sort into another bin*, recorded on a pair of UFACTORY xArm7 arms with parallel grippers, driven by one operator with a Meta Quest headset, kept as they were recorded: 4 color cameras at their recorded resolution, the overhead camera's infrared stereo pair and depth from every camera, all stored without loss, together with the arms' joint states, joint commands and tool poses, a wrist force/torque sensor, the operator's subtask marks and the episode outcome, on one clock, in the [LeRobot](https://github.com/huggingface/lerobot) v3.0 format. Collected by Paddy Data on its own rig.
> **Everything as recorded.** This release keeps all the data the rig captured at recording time, at full quality: every camera at the resolution it recorded, the overhead camera's infrared stereo pair, and depth from every camera, both aligned to color and as the camera's own image. Every frame is stored without loss. The recording ran on a workstation with an NVIDIA GeForce RTX 5090 graphics card. Future datasets may be reduced, with fewer streams, lower resolutions or lossy compression, to suit the recording hardware or a customer's requirements.
The lighter [`PaddyDataCollection/cable_sorting_sample`](https://huggingface.co/datasets/PaddyDataCollection/cable_sorting_sample) has all 25 episodes of these sessions with three cameras in standard-quality video and no depth (3.24 GB). This release's episodes 0, 1, 2 and 3 are its episodes 10, 13, 14 and 17.
## At a glance
- **Episodes:** 4 (3 success, 1 failure), 4,143 rows at 30 fps, 2.3 minutes
- **Task:** Pick cable from one bin and sort into another bin (one task, three operator-marked subtasks)
- **Robot:** two UFACTORY xArm7 (7 joints each), xArm Gripper G2 (0 to 84 mm), six-axis force/torque sensor on the right wrist
- **Teleoperation:** Meta Quest, one hand controller per arm
- **Cameras:** 4 Intel RealSense (`camera_bottom` scene, `camera_top` overhead, `camera_left` wrist, `camera_right` wrist), color at 30 fps: `camera_top` 1920x1080, the others 1280x720, lossless H.264 (High 4:4:4 Predictive)
- **Stereo:** the overhead camera's two infrared imagers, 1280x720, rectified, 49.9 mm baseline, emitter off, with a measured calibration (Sampson error P95 0.57 px on these episodes)
- **Depth:** every camera, 16-bit millimetres, aligned to its color image and as the camera's own 640x360 image, FFV1 lossless; also as LeRobot depth features
- **Recorded:** 2026-11-13 (UTC-5), 2 sessions, 2 scene layouts, 1 operator
- **Action:** absolute joint positions plus gripper command, 16 values per row
- **Size:** 19.38 GB (about 8.4 GB per minute), LeRobot v3.0
- **License:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/), attribution: Paddy Data
## Get it and load it
```bash
pip install -U "huggingface_hub[cli]"
hf download PaddyDataCollection/cable_sorting_sample_full --repo-type dataset --local-dir cable_sorting_sample_full
```
```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("PaddyDataCollection/cable_sorting_sample_full", root="cable_sorting_sample_full")
print(ds.num_episodes, ds.num_frames, ds.fps)
frame = ds[0]
print(frame["task"])
print(frame["observation.state"].shape, frame["action"].shape)
print(frame["observation.images.camera_top"].shape)
```
Output (LeRobot 0.6.1):
```text
4 4143 30
Pick cable from one bin and sort into another bin
torch.Size([16]) torch.Size([16])
torch.Size([3, 1080, 1920])
```
Videos decode with either of LeRobot's video backends (`pyav`, `torchcodec`). Each episode is one data
file (`data/chunk-000/file-NNN.parquet`) and one video file per camera, `NNN` being the episode index.
The color and stereo videos use H.264's lossless High 4:4:4 Predictive profile: FFmpeg, PyAV,
torchcodec and VLC play it, some players (QuickTime, web browsers) do not.
The depth files under `depth/` are not LeRobot features; read them with any FFmpeg-based reader, for
example PyAV:
```python
import av
import numpy as np
# depth of camera_top for episode 0, one frame per row of the episode
path = "cable_sorting_sample_full/depth/observation.depth.camera_top/chunk-000/file-000.mkv"
with av.open(path) as container:
depth_mm = np.stack([f.to_ndarray() for f in container.decode(video=0)])
print(depth_mm.shape, depth_mm.dtype) # uint16 millimetres, 0 = no measurement
```
Output:
```text
(1147, 1080, 1920) uint16
```
## What each row holds
One row per camera frame time, 30 per second. Each row holds the newest sample of every stream at or
before the row's time; nothing is interpolated.
| Group | Features (dimension) | Content |
|---|---|---|
| Cameras | `observation.images.camera_bottom`, `observation.images.camera_top`, `observation.images.camera_left`, `observation.images.camera_right` | color video, lossless H.264 (High 4:4:4 Predictive), `camera_top` 1920x1080, the others 1280x720 at 30 fps, one frame per row |
| Stereo pair | `observation.images.camera_top_stereo_left`, `observation.images.camera_top_stereo_right` | the overhead camera's two infrared imagers, rectified, emitter off, 1280x720; gray in three equal channels, lossless H.264 (High 4:4:4 Predictive) |
| Depth as LeRobot features | `observation.images.camera_bottom_depth`, `observation.images.camera_top_depth`, `observation.images.camera_left_depth`, `observation.images.camera_right_depth` | each camera's aligned depth in LeRobot's depth format: a 12-bit log code, HEVC (gray12le), about 1.3 mm per code at 0.3 to 1 m; the exact millimetres are in `depth/` |
| Joint positions | `observation.state` (16) | per arm joint1..joint7 (rad), then the gripper opening (0 closed .. 1 open) |
| Joint velocity and effort | `observation.velocity` (14), `observation.effort` (14) | per arm joint1..joint7, as the arm's SDK reports them (rad/s; N m). Effort is the SDK's estimate from motor current, not a torque sensor, and has not been validated |
| TCP pose | `observation.eef_pose.left` (7), `observation.eef_rotvec.left` (3), `observation.eef_rot6d.left` (6), `observation.eef_pose.right` (7), `observation.eef_rotvec.right` (3), `observation.eef_rot6d.right` (6), `observation.eef_state` (16) | measured tool pose in each arm's base frame (m, quaternion xyzw; 6D rotation; rotation vector), forward kinematics of the included URDF at the row's joints; eef_state packs both arms and grippers |
| Force/torque | `observation.force_torque.right` (6), `paddy.force_torque_raw.right` (6), `paddy.force_torque_valid` | right wrist, fx fy fz (N) and tx ty tz (N m) in the flange frame; compensated, raw, and a fresh-reading flag |
| Action | `action` (16) | per arm the 7 joint commands (rad, absolute), then the gripper command (0 closed .. 1 open) |
| Teleoperation target and sent pose | `paddy.command_sent.left` (7), `paddy.command_sent.right` (7), `paddy.target_pose.left` (7), `paddy.target_pose.right` (7), `paddy.target_rotvec.left` (3), `paddy.target_rotvec.right` (3), `paddy.target_rot6d.left` (6), `paddy.target_rot6d.right` (6) | the headset's TCP target per arm, and the pose the arm was commanded to |
| Per-arm flags | `paddy.action_valid` (2), `paddy.action_issued` (2), `paddy.engaged` (2), `paddy.scaling` (2), `paddy.moving` (2), `paddy.control_source` (2) | live command, command issued, deadman held, arm moving, who drove it (1 = human), motion scaling |
| Wrist camera poses | `paddy.camera_pose.camera_left` (7), `paddy.camera_pose.camera_right` (7) | each wrist camera's optical frame in its arm's base frame at the row (m, quaternion xyzw) |
| Time | `timestamp`, `paddy.utc_s`, `paddy.frame_age.camera_bottom`, `paddy.frame_age.camera_top`, `paddy.frame_age.camera_left`, `paddy.frame_age.camera_right`, `paddy.frame_age.camera_top_stereo_left`, `paddy.frame_age.camera_top_stereo_right` | seconds from episode start; UTC epoch seconds; age of each camera's frame at the row |
| Labels and indices | `paddy.subtask_index`, `frame_index`, `episode_index`, `index`, `task_index` | task string index, operator-marked subtask (-1 = none), LeRobot indices |
`meta/modality.json` names the per-arm slices of `observation.state` and `action`. Units, frames and the
timing rule of every column are written out in `meta/paddy/info.json` (`conventions`).
## Before you train on it
- **Action.** `action` is the joint command each arm received: per arm 7 absolute joint positions (radians,
the joint order of `observation.state`), then the gripper command (0 closed, 1 open). The arms ran in
servo mode: the control loop streamed a joint setpoint to each arm 200 times a second, and row k
holds the last setpoint sent at or before row k's time. State and action are in the same joint space.
Before an arm's first command the row holds its measured joints. The headset's Cartesian target is
beside it in `paddy.target_pose.<side>`.
The robot streams are on the 30 fps rows as in the sample; the 200 Hz setpoint stream at its own
rate and the headset's raw poses are not in this release.
- **`paddy.action_valid`** (one flag per arm, `[left, right]`) is true on rows where that arm had a live
teleoperation command: a fresh target from the headset, the arm enabled, no rejected targets. It is not a
measure of tracking quality. The arms take turns in this task (pick with the left arm, hand over, place
with the right), so each arm is commanded on 39 to 63 % of an episode's
rows. Mask or weight the idle arm's rows if your method needs it.
- **Subtasks.** The operator marked three subtasks live with a foot pedal:
- `0`: Pick up a cable with the left arm and use the right arm for a better camera view
- `1`: Pass the cable from the left arm to the right arm
- `2`: Place the green cables into the left bin and the black cables into the right bin
`paddy.subtask_index` gives the open subtask per row and -1 outside the marks (2.9 % of
rows, all of them before the first or after the last mark of an episode). Each episode's spans and per-subtask outcomes are in its
sidecar, `meta/paddy/episodes/<trial>.json`, under `conversion.subtasks`. The marks of one episode were corrected in review after recording: in episode 2, which failed, the step that failed was labelled `failure` from the operator's note and the video.
- **Failures.** 1 of 4 episodes (25 %, 27.1 % of the rows) is a failure, labelled `failure` in `paddy/outcome` (`meta/episodes`). The subtask that failed is labelled
`failure` and no later subtask is marked:
- episode 2 (37 s): failed at subtask 1 (handover); operator's note: "dropped cable"
Filter on `paddy/outcome` if you train on successes only. These 4 episodes were chosen from the 25 of the sample to show both scenes and both outcomes.
- **Scenes.** Two bin layouts, named per episode in `paddy/scene_id`: `Paddy_Office` (episodes 0-1); `Paddy_Office_binwall_2_lowlight` (episodes 2-3). In the first the bins
hang lower on the bin wall; in the second they hang higher and the room light is reduced (label `low_light`
in `paddy/labels`).
- **Timing.** All streams share one clock, the recording computer's, stored per row as UTC epoch seconds
in `paddy.utc_s`. Joint states arrive at 100 Hz and force/torque readings at about
100 Hz; force/torque is fresh on 100.00 % of rows (`paddy.force_torque_valid`). The
cameras' frames of a row are a median 7.2 to 14.1 ms apart per episode (the two
cameras furthest apart; 10.6 ms over the episodes). Each arm follows its command
about 116 ms later (110 to 123 ms per episode and arm),
so the state lags the action by about 3 rows. Where a camera
missed a frame the previous one is repeated; `paddy.frame_age.<camera>` gives the age of each
camera's frame at the row (at most 49 ms in this dataset).
- **Episode boundaries.** An episode covers the time in which every included stream has data: it starts
when the last stream has started and ends when the first one stops. No row inside an episode is
dropped and no episode was paused. `meta/paddy/keep_ranges.json` lists per episode the row
ranges in which an arm moves (`paddy.moving`), if you want to drop idle rows at the ends.
- **Ratings and ids.** `paddy/rating` is the operator's 1 to 5 judgement given at the stop, 0 where
none was given (2 of 4 episodes rated). `paddy/episode_uuid` identifies each recording and is the
same in both releases.
## Cameras and calibration
| Camera | Model | Role | Resolution | Intrinsics | Extrinsics (ChArUco board) |
|---|---|---|---|---|---|
| `camera_bottom` | Intel RealSense D405 | scene | 1280x720, as recorded | factory K and distortion | 2026-11-12, 33 views, reprojection RMS 1.23 px, 0.34 mm |
| `camera_top` | Intel RealSense D435 | overhead, primary | 1920x1080, as recorded | factory K, distortion fitted with a ChArUco board | 2026-11-12, 31 views, reprojection RMS 0.55 px, 0.43 mm |
| `camera_left` | Intel RealSense D405 | wrist | 1280x720, as recorded | factory K and distortion | 2026-11-12, 17 views, reprojection RMS 1.01 px, 0.85 mm |
| `camera_right` | Intel RealSense D405 | wrist | 1280x720, as recorded | factory K and distortion | 2026-11-12, 16 views, reprojection RMS 0.88 px, 0.73 mm |
- Intrinsics (`K`, `D`, plumb_bob) in `meta/cameras.json` describe the stored video, which is not
undistorted.
- Extrinsics: `meta/cameras.json` gives each camera's color optical frame in each arm's base frame
(`extrinsics.<base>.base_T_cam`, metres). The wrist cameras' pose at every row is in
`paddy.camera_pose.<camera>`, from the joints through the URDF and the measured mount.
The calibration records note that `camera_bottom`, `camera_left` and `camera_right` missed the calibration tool's reprojection target (0.5 to 0.7 px).
- Robot description: `meta/paddy/robot_description.urdf` holds both arms, grippers, the force/torque
sensor and the cameras in one tree, with the joint origins fitted to camera views of a ChArUco board.
Mesh files are not included; kinematics and inertials are complete without them.
## What is stored
The cameras deliver color and infrared frames JPEG-compressed and depth as 16-bit millimetres; this release
stores every one of those frames without further loss. Color and infrared are decoded, converted to 4:2:0
YUV (BT.601, limited range) as every video here is, and encoded with H.264 at qp 0, which is lossless; depth
is FFV1.
| Stream | Feature | Size | Encoding | Lossless | Bytes |
|---|---|---|---|---|---|
| Color | `observation.images.camera_bottom` | 1280x720 | lossless H.264 (High 4:4:4 Predictive) | yes, after the 4:2:0 conversion | 1.76 GB |
| Color | `observation.images.camera_top` | 1920x1080 | lossless H.264 (High 4:4:4 Predictive) | yes, after the 4:2:0 conversion | 3.31 GB |
| Color | `observation.images.camera_left` | 1280x720 | lossless H.264 (High 4:4:4 Predictive) | yes, after the 4:2:0 conversion | 1.78 GB |
| Color | `observation.images.camera_right` | 1280x720 | lossless H.264 (High 4:4:4 Predictive) | yes, after the 4:2:0 conversion | 1.83 GB |
| Infrared stereo pair | `observation.images.camera_top_stereo_left` | 1280x720 | lossless H.264 (High 4:4:4 Predictive) | yes, after the 4:2:0 conversion | 0.80 GB |
| Infrared stereo pair | `observation.images.camera_top_stereo_right` | 1280x720 | lossless H.264 (High 4:4:4 Predictive) | yes, after the 4:2:0 conversion | 0.80 GB |
| Depth, aligned to color | `observation.depth.camera_bottom` | 1280x720 | FFV1, 16-bit mm | yes | 0.53 GB |
| Depth, aligned to color | `observation.depth.camera_top` | 1920x1080 | FFV1, 16-bit mm | yes | 0.34 GB |
| Depth, aligned to color | `observation.depth.camera_left` | 1280x720 | FFV1, 16-bit mm | yes | 0.45 GB |
| Depth, aligned to color | `observation.depth.camera_right` | 1280x720 | FFV1, 16-bit mm | yes | 0.53 GB |
| Depth, the camera's own image | `observation.depth_raw.camera_bottom` | 640x360 | FFV1, 16-bit mm | yes | 0.38 GB |
| Depth, the camera's own image | `observation.depth_raw.camera_top` | 640x360 | FFV1, 16-bit mm | yes | 0.29 GB |
| Depth, the camera's own image | `observation.depth_raw.camera_left` | 640x360 | FFV1, 16-bit mm | yes | 0.33 GB |
| Depth, the camera's own image | `observation.depth_raw.camera_right` | 640x360 | FFV1, 16-bit mm | yes | 0.37 GB |
| Depth, LeRobot feature | `observation.images.camera_bottom_depth` | 1280x720 | HEVC, 12-bit log code | the code is stored losslessly; the code rounds depth to 1.25 to 1.48 mm at 0.3 to 1 m | 1.51 GB |
| Depth, LeRobot feature | `observation.images.camera_top_depth` | 1920x1080 | HEVC, 12-bit log code | the code is stored losslessly; the code rounds depth to 1.25 to 1.48 mm at 0.3 to 1 m | 1.55 GB |
| Depth, LeRobot feature | `observation.images.camera_left_depth` | 1280x720 | HEVC, 12-bit log code | the code is stored losslessly; the code rounds depth to 1.25 to 1.48 mm at 0.3 to 1 m | 1.32 GB |
| Depth, LeRobot feature | `observation.images.camera_right_depth` | 1280x720 | HEVC, 12-bit log code | the code is stored losslessly; the code rounds depth to 1.25 to 1.48 mm at 0.3 to 1 m | 1.50 GB |
## Depth
- Aligned: `depth/observation.depth.<camera>/chunk-000/file-NNN.mkv` (16 files), 16-bit millimetres
(`gray16le`), one frame per row, on its camera's color pixel grid with the same `K`.
- Raw: `depth/observation.depth_raw.<camera>/chunk-000/file-NNN.mkv` (16 files), the camera's own
640x360 depth image, not aligned, one frame per row; its intrinsics are `depth` and its pose relative to
the color camera `color_T_depth` in `meta/cameras.json`.
- LeRobot depth features `observation.images.<camera>_depth` hold the aligned depth in LeRobot's format,
a 12-bit log code that LeRobot loads as millimetres: one code is 1.25 mm at 0.3 m and
1.48 mm at 1 m. Use the FFV1 files for exact values.
- 0 means no measurement (the LeRobot features load it as 10 mm). Rare far values are
sensor noise: clip depth to your working range.
- 2 of the 16,572 aligned depth frames are all zero because no depth frame had arrived for
that row (episode 0 `camera_top` 1; episode 2 `camera_bottom` 1); the color frame is there.
- `meta/paddy/info.json` (`depth`) describes the layout.
## Stereo pair
`observation.images.camera_top_stereo_left` and `_right` are the overhead camera's two infrared imagers,
rectified by the camera, with the emitter off, paired by capture time and stored like color (the gray
image in three equal channels).
The pair has a measured calibration, made on 2026-11-17 with a ChArUco board (9x7 squares of 60 mm)
with the camera where it was for these recordings: 100 views of the pair, 80 fitted and 20 held out.
Sampson error is the epipolar error of matched points, in pixels.
| | Measured calibration | Camera's factory calibration |
|---|---|---|
| Reprojection RMS, fitted views | 0.43 px | |
| Sampson error on the held-out board views, P95 | 0.17 px | 1.16 px |
| Sampson error on this release's stereo videos, P95 (SIFT matches, 16 pairs per episode) | 0.57 px (0.55 to 0.58 per episode) | 1.59 px |
| Baseline | 49.90 mm | 49.98 mm |
Matched points in the stored pair lie about 1.6 px apart vertically, which the camera's rectification
leaves and the measured calibration describes. On the videos, a fundamental matrix fitted to the same
matches does no better than the calibration (P95 0.57 to 0.61 px), so the 0.57 px is the matching
noise. The baseline's scale rests on the board's printed 60 mm squares.
Under `cameras.camera_top.stereo` in `meta/cameras.json`: each imager's `K`, `D` and `P` (measured; the
camera's own values under `factory`), `left_T_right`, the baseline, the pose of the left imager
relative to the color camera, and under `calibration` R, T, E, F and the errors.
`meta/paddy/stereo_calibration_camera_top.json` is the full calibration with its per-view errors, and
`meta/paddy/stereo_calibration_camera_top.yaml` the same for OpenCV's `FileStorage`.
The measured calibration was added on 2026-11-17, after the first upload; the videos and data did not
change. The earlier revision, with the factory calibration only, is in this repository's history.
## Difference from `PaddyDataCollection/cable_sorting_sample`
[`PaddyDataCollection/cable_sorting_sample`](https://huggingface.co/datasets/PaddyDataCollection/cable_sorting_sample) holds all 25 episodes with
`camera_top`, `camera_left` and `camera_right` only, in standard-quality H.264 (the overhead camera scaled
to 1280x720), and no depth or stereo pair. This release's episodes are its
episodes 10, 13, 14 and 17. Episode 2 here starts 79 rows (2.6 s) later than episode 14 there, because `camera_bottom` started last. The other 3 have
the same rows in both (every shared column equal but the dataset-wide `index` and `episode_index`).
## Files
- `data/`, `videos/`, `depth/`, `meta/info.json`, `meta/stats.json`, `meta/tasks.parquet`,
`meta/episodes/`: LeRobot v3.0.
- `meta/cameras.json`: intrinsics, extrinsics, calibration records, camera roles, the stereo calibration.
- `meta/paddy/stereo_calibration_camera_top.json` and `.yaml`: the stereo pair's measured calibration.
- `meta/paddy/info.json`: conventions for every column (units, frames, timing, masks), rates.
- `meta/paddy/episodes/<trial>.json`: one sidecar per episode with the recorder's metadata (labels,
notes, subtasks, a snapshot of the rig configuration) and the conversion report (rates, stream gaps,
timing). In these files, hardware serial numbers, network addresses and interface names are
replaced with `<removed>`, home folders with `/home/user`, and the operator's name with
`operator_01`; `rig_hash` is the hash of the configuration as recorded.
- `meta/paddy/logs/<session>/<trial>/`: the recorder's logs and each raw recording's `metadata.yaml`.
- `meta/paddy/keep_ranges.json`, `meta/paddy/robot_description.urdf`.
- `manifest.json` (episodes, summary, every file with its size and sha256) and `SHA256SUMS`: check a
copy with `sha256sum -c SHA256SUMS` in its root.
## People in view
- **Declared:** others (other people may appear in the cameras' view)
Faces are not blurred.
## License and attribution
Creative Commons Attribution 4.0 International (CC BY 4.0): https://creativecommons.org/licenses/by/4.0/. You may share and adapt the data, also commercially, if you give
appropriate credit. Attribution: Paddy Data, https://huggingface.co/datasets/PaddyDataCollection/cable_sorting_sample_full.
## Contact
Paddy Data: https://paddydata.ai
<details>
<summary>meta/info.json</summary>
```json
{
"codebase_version": "v3.0",
"robot_type": "dual_xarm7",
"total_episodes": 4,
"total_frames": 4143,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"fps": 30,
"splits": {
"train": "0:4"
},
"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",
"features": {
"observation.state": {
"dtype": "float32",
"shape": [
16
],
"names": [
"left_joint1",
"left_joint2",
"left_joint3",
"left_joint4",
"left_joint5",
"left_joint6",
"left_joint7",
"left_gripper",
"right_joint1",
"right_joint2",
"right_joint3",
"right_joint4",
"right_joint5",
"right_joint6",
"right_joint7",
"right_gripper"
],
"info": {
"unit": "rad; gripper 0 closed .. 1 open",
"order": "per arm joint1..jointN (URDF order, base to wrist), then the gripper",
"conventions": "meta/paddy/info.json conventions.observation_state"
}
},
"observation.velocity": {
"dtype": "float32",
"shape": [
14
],
"names": [
"left_joint1",
"left_joint2",
"left_joint3",
"left_joint4",
"left_joint5",
"left_joint6",
"left_joint7",
"right_joint1",
"right_joint2",
"right_joint3",
"right_joint4",
"right_joint5",
"right_joint6",
"right_joint7"
],
"info": {
"unit": "rad/s",
"source": "sdk",
"meaning": "joint velocities the xArm SDK reports (get_joint_states), as the follower published them with each joint-state sample (100 Hz), zero-order held like observation.state",
"conventions": "meta/paddy/info.json conventions.velocity",
"order": "per arm joint1..jointN (URDF order, base to wrist), as observation.state; no gripper"
}
},
"observation.effort": {
"dtype": "float32",
"shape": [
14
],
"names": [
"left_joint1",
"left_joint2",
"left_joint3",
"left_joint4",
"left_joint5",
"left_joint6",
"left_joint7",
"right_joint1",
"right_joint2",
"right_joint3",
"right_joint4",
"right_joint5",
"right_joint6",
"right_joint7"
],
"info": {
"unit": "N m",
"source": "sdk_estimate",
"validated": false,
"meaning": "the xArm SDK's joint torque estimate (get_joint_states) from motor current and the arm's own dynamic model; the arms have no joint torque sensors. Sign and magnitude not yet checked on the rig",
"conventions": "meta/paddy/info.json conventions.effort",
"order": "per arm joint1..jointN (URDF order, base to wrist), as observation.state; no gripper"
}
},
"observation.eef_pose.left": {
"dtype": "float32",
"shape": [
7
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw"
],
"info": {
"unit": "m",
"rotation": "quaternion xyzw",
"frame": "Left_link_base"
}
},
"observation.eef_rotvec.left": {
"dtype": "float32",
"shape": [
3
],
"names": [
"rx",
"ry",
"rz"
],
"info": {
"unit": "rad"
}
},
"observation.eef_rot6d.left": {
"dtype": "float32",
"shape": [
6
],
"names": [
"r00",
"r10",
"r20",
"r01",
"r11",
"r21"
],
"info": {
"unit": "1",
"rotation": "6D: the first two columns of the rotation matrix, column-major [r00, r10, r20, r01, r11, r21] (Zhou et al. 2019); zeros where the quaternion is zero",
"conventions": "meta/paddy/info.json conventions.rot6d",
"of": "observation.eef_pose.left",
"frame": "Left_link_base"
}
},
"observation.eef_pose.right": {
"dtype": "float32",
"shape": [
7
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw"
],
"info": {
"unit": "m",
"rotation": "quaternion xyzw",
"frame": "Right_link_base"
}
},
"observation.eef_rotvec.right": {
"dtype": "float32",
"shape": [
3
],
"names": [
"rx",
"ry",
"rz"
],
"info": {
"unit": "rad"
}
},
"observation.eef_rot6d.right": {
"dtype": "float32",
"shape": [
6
],
"names": [
"r00",
"r10",
"r20",
"r01",
"r11",
"r21"
],
"info": {
"unit": "1",
"rotation": "6D: the first two columns of the rotation matrix, column-major [r00, r10, r20, r01, r11, r21] (Zhou et al. 2019); zeros where the quaternion is zero",
"conventions": "meta/paddy/info.json conventions.rot6d",
"of": "observation.eef_pose.right",
"frame": "Right_link_base"
}
},
"observation.eef_state": {
"dtype": "float32",
"shape": [
16
],
"names": [
"left_x",
"left_y",
"left_z",
"left_qx",
"left_qy",
"left_qz",
"left_qw",
"left_gripper",
"right_x",
"right_y",
"right_z",
"right_qx",
"right_qy",
"right_qz",
"right_qw",
"right_gripper"
],
"info": {
"unit": "m; gripper 0 closed .. 1 open",
"rotation": "quaternion xyzw, sign-continuous per episode",
"frame": {
"left": "Left_link_base",
"right": "Right_link_base"
},
"layout_per_arm": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw",
"gripper"
],
"meaning": "observation.eef_pose.<side> and observation.state's gripper of every arm in one vector: the measured TCP pose and gripper opening in a tcp action's layout, the state of an end-effector policy; gripper 0 for an arm without one",
"conventions": "meta/paddy/info.json conventions.eef_state"
}
},
"action": {
"dtype": "float32",
"shape": [
16
],
"names": [
"left_joint1",
"left_joint2",
"left_joint3",
"left_joint4",
"left_joint5",
"left_joint6",
"left_joint7",
"left_gripper_cmd",
"right_joint1",
"right_joint2",
"right_joint3",
"right_joint4",
"right_joint5",
"right_joint6",
"right_joint7",
"right_gripper_cmd"
],
"info": {
"unit": "rad; gripper 0 closed .. 1 open",
"space": "joints",
"order": "per arm joint1..jointN (URDF order, base to wrist, as in observation.state), then the gripper command",
"joints": "absolute joint positions (not a delta)",
"timing": "row k holds the last joint command stamped at or before t_k (zero-order hold), the measured joints before the first one; without a recorded joint command, the next row's measured joints (meta/paddy/info.json conventions.action.joints_source)",
"conventions": "meta/paddy/info.json conventions.action"
}
},
"paddy.action_valid": {
"dtype": "bool",
"shape": [
2
],
"names": [
"left",
"right"
]
},
"paddy.action_issued": {
"dtype": "bool",
"shape": [
2
],
"names": [
"left",
"right"
]
},
"paddy.command_sent.left": {
"dtype": "float32",
"shape": [
7
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw"
],
"info": {
"unit": "m",
"rotation": "quaternion xyzw",
"frame": "Left_link_base"
}
},
"paddy.command_sent.right": {
"dtype": "float32",
"shape": [
7
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw"
],
"info": {
"unit": "m",
"rotation": "quaternion xyzw",
"frame": "Right_link_base"
}
},
"observation.force_torque.right": {
"dtype": "float32",
"shape": [
6
],
"names": [
"fx",
"fy",
"fz",
"tx",
"ty",
"tz"
],
"info": {
"unit": "N (fx fy fz), N m (tx ty tz)",
"frame": "Right_link_eef",
"timing": "row k holds the last reading stamped at or before t_k (zero-order hold); zeros where paddy.force_torque_valid is false",
"conventions": "meta/paddy/info.json conventions.force_torque",
"reading": "after the control box's load and offset compensation"
}
},
"paddy.force_torque_raw.right": {
"dtype": "float32",
"shape": [
6
],
"names": [
"fx",
"fy",
"fz",
"tx",
"ty",
"tz"
],
"info": {
"unit": "N (fx fy fz), N m (tx ty tz)",
"frame": "Right_link_eef",
"timing": "row k holds the last reading stamped at or before t_k (zero-order hold); zeros where paddy.force_torque_valid is false",
"conventions": "meta/paddy/info.json conventions.force_torque",
"reading": "unprocessed sensor reading, no compensation"
}
},
"paddy.force_torque_valid": {
"dtype": "bool",
"shape": [
1
],
"names": [
"right"
]
},
"paddy.target_pose.left": {
"dtype": "float32",
"shape": [
7
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw"
],
"info": {
"unit": "m",
"rotation": "quaternion xyzw, sign-continuous per episode",
"frame": "Left_link_base",
"pose": "absolute TCP target in the arm base frame (not a delta)",
"timing": "row k holds the last target stamped at or before t_k (zero-order hold)"
}
},
"paddy.target_pose.right": {
"dtype": "float32",
"shape": [
7
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw"
],
"info": {
"unit": "m",
"rotation": "quaternion xyzw, sign-continuous per episode",
"frame": "Right_link_base",
"pose": "absolute TCP target in the arm base frame (not a delta)",
"timing": "row k holds the last target stamped at or before t_k (zero-order hold)"
}
},
"paddy.engaged": {
"dtype": "bool",
"shape": [
2
],
"names": [
"left",
"right"
]
},
"paddy.scaling": {
"dtype": "float32",
"shape": [
2
],
"names": [
"left",
"right"
]
},
"paddy.moving": {
"dtype": "bool",
"shape": [
2
],
"names": [
"left",
"right"
],
"info": {
"meaning": "the arm moves at the row: TCP, an arm joint or the gripper over a window centred on the row",
"rule": "meta/paddy/info.json motion"
}
},
"paddy.control_source": {
"dtype": "int8",
"shape": [
2
],
"names": [
"left",
"right"
],
"info": {
"meaning": "who drove the arm at the row",
"values": {
"0": "nobody",
"1": "human",
"2": "policy"
}
}
},
"paddy.target_rotvec.left": {
"dtype": "float32",
"shape": [
3
],
"names": [
"rx",
"ry",
"rz"
],
"info": {
"unit": "rad"
}
},
"paddy.target_rotvec.right": {
"dtype": "float32",
"shape": [
3
],
"names": [
"rx",
"ry",
"rz"
],
"info": {
"unit": "rad"
}
},
"paddy.target_rot6d.left": {
"dtype": "float32",
"shape": [
6
],
"names": [
"r00",
"r10",
"r20",
"r01",
"r11",
"r21"
],
"info": {
"unit": "1",
"rotation": "6D: the first two columns of the rotation matrix, column-major [r00, r10, r20, r01, r11, r21] (Zhou et al. 2019); zeros where the quaternion is zero",
"conventions": "meta/paddy/info.json conventions.rot6d",
"of": "paddy.target_pose.left",
"frame": "Left_link_base",
"not_a_policy_output": true
}
},
"paddy.target_rot6d.right": {
"dtype": "float32",
"shape": [
6
],
"names": [
"r00",
"r10",
"r20",
"r01",
"r11",
"r21"
],
"info": {
"unit": "1",
"rotation": "6D: the first two columns of the rotation matrix, column-major [r00, r10, r20, r01, r11, r21] (Zhou et al. 2019); zeros where the quaternion is zero",
"conventions": "meta/paddy/info.json conventions.rot6d",
"of": "paddy.target_pose.right",
"frame": "Right_link_base",
"not_a_policy_output": true
}
},
"paddy.subtask_index": {
"dtype": "int64",
"shape": [
1
],
"names": null,
"info": {
"meaning": "index of the operator-marked subtask open at the row, -1 = none",
"names": "meta/paddy/episodes/<trial>.json conversion.subtasks"
}
},
"timestamp": {
"dtype": "float32",
"shape": [
1
],
"names": null
},
"paddy.utc_s": {
"dtype": "float64",
"shape": [
1
],
"names": null,
"info": {
"unit": "s",
"clock": "UTC epoch seconds, ROS header.stamp of the recorder host"
}
},
"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.camera_bottom": {
"dtype": "video",
"shape": [
3,
720,
1280
],
"names": [
"channel",
"height",
"width"
],
"info": {
"video.height": 720,
"video.width": 1280,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"video.channels": 3,
"video.g": 10,
"video.crf": 0,
"video.preset": "medium",
"video.is_depth_map": false,
"has_audio": false
}
},
"observation.images.camera_top": {
"dtype": "video",
"shape": [
3,
1080,
1920
],
"names": [
"channel",
"height",
"width"
],
"info": {
"video.height": 1080,
"video.width": 1920,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"video.channels": 3,
"video.g": 10,
"video.crf": 0,
"video.preset": "medium",
"video.is_depth_map": false,
"has_audio": false
}
},
"observation.images.camera_left": {
"dtype": "video",
"shape": [
3,
720,
1280
],
"names": [
"channel",
"height",
"width"
],
"info": {
"video.height": 720,
"video.width": 1280,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"video.channels": 3,
"video.g": 10,
"video.crf": 0,
"video.preset": "medium",
"video.is_depth_map": false,
"has_audio": false
}
},
"observation.images.camera_right": {
"dtype": "video",
"shape": [
3,
720,
1280
],
"names": [
"channel",
"height",
"width"
],
"info": {
"video.height": 720,
"video.width": 1280,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"video.channels": 3,
"video.g": 10,
"video.crf": 0,
"video.preset": "medium",
"video.is_depth_map": false,
"has_audio": false
}
},
"observation.images.camera_top_stereo_left": {
"dtype": "video",
"shape": [
3,
720,
1280
],
"names": [
"channel",
"height",
"width"
],
"info": {
"video.height": 720,
"video.width": 1280,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"video.channels": 3,
"video.g": 10,
"video.crf": 0,
"video.preset": "medium",
"video.is_depth_map": false,
"has_audio": false,
"paddy.stereo_camera": "camera_top",
"paddy.stereo_side": "left",
"paddy.stereo_stream": "infra1",
"paddy.source_channels": 1,
"paddy.spectrum": "infrared"
}
},
"observation.images.camera_top_stereo_right": {
"dtype": "video",
"shape": [
3,
720,
1280
],
"names": [
"channel",
"height",
"width"
],
"info": {
"video.height": 720,
"video.width": 1280,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"video.channels": 3,
"video.g": 10,
"video.crf": 0,
"video.preset": "medium",
"video.is_depth_map": false,
"has_audio": false,
"paddy.stereo_camera": "camera_top",
"paddy.stereo_side": "right",
"paddy.stereo_stream": "infra2",
"paddy.source_channels": 1,
"paddy.spectrum": "infrared"
}
},
"paddy.frame_age.camera_bottom": {
"dtype": "float32",
"shape": [
1
],
"names": null,
"info": {
"unit": "s"
}
},
"paddy.frame_age.camera_top": {
"dtype": "float32",
"shape": [
1
],
"names": null,
"info": {
"unit": "s"
}
},
"paddy.frame_age.camera_left": {
"dtype": "float32",
"shape": [
1
],
"names": null,
"info": {
"unit": "s"
}
},
"paddy.frame_age.camera_right": {
"dtype": "float32",
"shape": [
1
],
"names": null,
"info": {
"unit": "s"
}
},
"paddy.frame_age.camera_top_stereo_left": {
"dtype": "float32",
"shape": [
1
],
"names": null,
"info": {
"unit": "s",
"of": "observation.images.camera_top_stereo_left",
"pair": "equal on both views"
}
},
"paddy.frame_age.camera_top_stereo_right": {
"dtype": "float32",
"shape": [
1
],
"names": null,
"info": {
"unit": "s",
"of": "observation.images.camera_top_stereo_right",
"pair": "equal on both views"
}
},
"paddy.camera_pose.camera_left": {
"dtype": "float32",
"shape": [
7
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw"
],
"info": {
"unit": "m",
"frame": "Left_link_base",
"camera_frame": "Left_camera_color_optical_frame",
"time": "row"
}
},
"paddy.camera_pose.camera_right": {
"dtype": "float32",
"shape": [
7
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw"
],
"info": {
"unit": "m",
"frame": "Right_link_base",
"camera_frame": "Right_camera_color_optical_frame",
"time": "row"
}
},
"observation.images.camera_bottom_depth": {
"dtype": "video",
"shape": [
720,
1280,
1
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 720,
"video.width": 1280,
"video.codec": "hevc",
"video.pix_fmt": "gray12le",
"video.fps": 30,
"video.channels": 1,
"has_audio": false,
"video.g": 2,
"video.crf": 30,
"video.preset": null,
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": {
"x265-params": "lossless=1:open-gop=0:log-level=error"
},
"video.depth_min": 0.01,
"video.depth_max": 10.0,
"video.shift": 3.5,
"video.use_log": true,
"is_depth_map": true,
"depth_unit": "mm"
}
},
"observation.images.camera_top_depth": {
"dtype": "video",
"shape": [
1080,
1920,
1
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 1080,
"video.width": 1920,
"video.codec": "hevc",
"video.pix_fmt": "gray12le",
"video.fps": 30,
"video.channels": 1,
"has_audio": false,
"video.g": 2,
"video.crf": 30,
"video.preset": null,
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": {
"x265-params": "lossless=1:open-gop=0:log-level=error"
},
"video.depth_min": 0.01,
"video.depth_max": 10.0,
"video.shift": 3.5,
"video.use_log": true,
"is_depth_map": true,
"depth_unit": "mm"
}
},
"observation.images.camera_left_depth": {
"dtype": "video",
"shape": [
720,
1280,
1
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 720,
"video.width": 1280,
"video.codec": "hevc",
"video.pix_fmt": "gray12le",
"video.fps": 30,
"video.channels": 1,
"has_audio": false,
"video.g": 2,
"video.crf": 30,
"video.preset": null,
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": {
"x265-params": "lossless=1:open-gop=0:log-level=error"
},
"video.depth_min": 0.01,
"video.depth_max": 10.0,
"video.shift": 3.5,
"video.use_log": true,
"is_depth_map": true,
"depth_unit": "mm"
}
},
"observation.images.camera_right_depth": {
"dtype": "video",
"shape": [
720,
1280,
1
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 720,
"video.width": 1280,
"video.codec": "hevc",
"video.pix_fmt": "gray12le",
"video.fps": 30,
"video.channels": 1,
"has_audio": false,
"video.g": 2,
"video.crf": 30,
"video.preset": null,
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": {
"x265-params": "lossless=1:open-gop=0:log-level=error"
},
"video.depth_min": 0.01,
"video.depth_max": 10.0,
"video.shift": 3.5,
"video.use_log": true,
"is_depth_map": true,
"depth_unit": "mm"
}
}
}
}
```
</details>
license: cc-by-4.0
pretty_name: "Cable sorting sample, full: every recorded stream, lossless (bimanual xArm7)"
task_categories:
- robotics
tags:
- LeRobot
- robotics
- teleoperation
- bimanual
- manipulation
- xarm
- depth
- stereo
- lossless
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files: data/*/*.parquet
# Cable sorting sample, full (lossless)
4 teleoperated episodes of one bimanual task, *Pick cable from one bin and sort into another bin*, recorded on a pair of UFACTORY xArm7 arms with parallel grippers, driven by one operator with a Meta Quest headset, kept as they were recorded: 4 color cameras at their recorded resolution, the overhead camera's infrared stereo pair and depth from every camera, all stored without loss, together with the arms' joint states, joint commands and tool poses, a wrist force/torque sensor, the operator's subtask marks and the episode outcome, on one clock, in the [LeRobot](https://github.com/huggingface/lerobot) v3.0 format. Collected by Paddy Data on its own rig.
> **Everything as recorded.** This release keeps all the data the rig captured at recording time, at full quality: every camera at the resolution it recorded, the overhead camera's infrared stereo pair, and depth from every camera, both aligned to color and as the camera's own image. Every frame is stored without loss. The recording ran on a workstation with an NVIDIA GeForce RTX 5090 graphics card. Future datasets may be reduced, with fewer streams, lower resolutions or lossy compression, to suit the recording hardware or a customer's requirements.
The lighter [`PaddyDataCollection/cable_sorting_sample`](https://huggingface.co/datasets/PaddyDataCollection/cable_sorting_sample) has all 25 episodes of these sessions with three cameras in standard-quality video and no depth (3.24 GB). This release's episodes 0, 1, 2 and 3 are its episodes 10, 13, 14 and 17.
## At a glance
- **Episodes:** 4 (3 success, 1 failure), 4,143 rows at 30 fps, 2.3 minutes
- **Task:** Pick cable from one bin and sort into another bin (one task, three operator-marked subtasks)
- **Robot:** two UFACTORY xArm7 (7 joints each), xArm Gripper G2 (0 to 84 mm), six-axis force/torque sensor on the right wrist
- **Teleoperation:** Meta Quest, one hand controller per arm
- **Cameras:** 4 Intel RealSense (`camera_bottom` scene, `camera_top` overhead, `camera_left` wrist, `camera_right` wrist), color at 30 fps: `camera_top` 1920x1080, the others 1280x720, lossless H.264 (High 4:4:4 Predictive)
- **Stereo:** the overhead camera's two infrared imagers, 1280x720, rectified, 49.9 mm baseline, emitter off, with a measured calibration (Sampson error P95 0.57 px on these episodes)
- **Depth:** every camera, 16-bit millimetres, aligned to its color image and as the camera's own 640x360 image, FFV1 lossless; also as LeRobot depth features
- **Recorded:** 2026-11-13 (UTC-5), 2 sessions, 2 scene layouts, 1 operator
- **Action:** absolute joint positions plus gripper command, 16 values per row
- **Size:** 19.38 GB (about 8.4 GB per minute), LeRobot v3.0
- **License:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/), attribution: Paddy Data
## Get it and load it
```bash
pip install -U "huggingface_hub[cli]"
hf download PaddyDataCollection/cable_sorting_sample_full --repo-type dataset --local-dir cable_sorting_sample_full
```
```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("PaddyDataCollection/cable_sorting_sample_full", root="cable_sorting_sample_full")
print(ds.num_episodes, ds.num_frames, ds.fps)
frame = ds[0]
print(frame["task"])
print(frame["observation.state"].shape, frame["action"].shape)
print(frame["observation.images.camera_top"].shape)
```
Output (LeRobot 0.6.1):
```text
4 4143 30
Pick cable from one bin and sort into another bin
torch.Size([16]) torch.Size([16])
torch.Size([3, 1080, 1920])
```
Videos decode with either of LeRobot's video backends (`pyav`, `torchcodec`). Each episode is one data
file (`data/chunk-000/file-NNN.parquet`) and one video file per camera, `NNN` being the episode index.
The color and stereo videos use H.264's lossless High 4:4:4 Predictive profile: FFmpeg, PyAV,
torchcodec and VLC play it, some players (QuickTime, web browsers) do not.
The depth files under `depth/` are not LeRobot features; read them with any FFmpeg-based reader, for
example PyAV:
```python
import av
import numpy as np
# depth of camera_top for episode 0, one frame per row of the episode
path = "cable_sorting_sample_full/depth/observation.depth.camera_top/chunk-000/file-000.mkv"
with av.open(path) as container:
depth_mm = np.stack([f.to_ndarray() for f in container.decode(video=0)])
print(depth_mm.shape, depth_mm.dtype) # uint16 millimetres, 0 = no measurement
```
Output:
```text
(1147, 1080, 1920) uint16
```
## What each row holds
One row per camera frame time, 30 per second. Each row holds the newest sample of every stream at or