task_categories:
- robotics
tags:
- LeRobot
- nero7
- block-sorting
- block-stacking
configs:
- config_name: default
data_files: data/*/*.parquet
This dataset was created using LeRobot.
Dataset Description
- Homepage: [More Information Needed]
- Paper: [More Information Needed]
- License: [More Information Needed]
Dataset Structure
{
"codebase_version": "v3.0",
"robot_type": "nero7_single_arm",
"total_episodes": 165,
"total_frames": 311338,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"fps": 30,
"splits": {
"train": "0:165"
},
"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.images.external": {
"dtype": "video",
"shape": [
360,
640,
3
],
"names": [
"height",
"width",
"channel"
],
"info": {
"video.height": 360,
"video.width": 640,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 30,
"video.channels": 3,
"has_audio": false
}
},
"observation.images.wrist": {
"dtype": "video",
"shape": [
300,
480,
3
],
"names": [
"height",
"width",
"channel"
],
"info": {
"video.height": 300,
"video.width": 480,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 30,
"video.channels": 3,
"has_audio": false
}
},
"observation.state": {
"dtype": "float32",
"shape": [
8
],
"names": [
"joint1",
"joint2",
"joint3",
"joint4",
"joint5",
"joint6",
"joint7",
"gripper"
]
},
"action": {
"dtype": "float32",
"shape": [
8
],
"names": [
"joint1",
"joint2",
"joint3",
"joint4",
"joint5",
"joint6",
"joint7",
"gripper"
]
},
"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
}
}
}
Citation
BibTeX:
[More Information Needed]
This is the successful-demonstration dataset (165 episodes).
Task and collection
Place the green and orange blocks in the cup, then stack the blue blocks. The active work area is the mat; spare objects outside it are not task targets. Five recording sessions were collected on 2026-10-24 and 2026-10-25, with more objects in later sessions. Exact object counts have not been annotated per episode. The cup position, cup appearance and object arrangement vary across sessions.
The default main / v3.0 revision contains 165 operator-labeled successful
demonstrations. The failures branch contains 10 operator-labeled failed
demonstrations, kept separately so they are not used as successful training examples.
Labels are the operator's recorded success/failure marks, not a new independent
visual certification of every trajectory.
| Session | Successful episodes | Failed episodes |
|---|---|---|
| 20261024_165431 | 45 | 3 |
| 20261024_175351 | 40 | 6 |
| 20261025_130101 | 32 | 0 |
| 20261025_140035 | 26 | 1 |
| 20261025_150001 | 22 | 0 |
Representation
- LeRobot v3.0, 30 frames/second, single AgileX Nero seven-axis follower arm.
observation.state: seven measured joint angles in radians, followed by measured total gripper opening in metres.action: seven leader joint targets in radians, followed by commanded total gripper opening in metres. Absolute joint targets are used.- External Gemini 336L RGB: 640 x 360; wrist Gemini 305 RGB: 480 x 300.
- Videos are H.264, without audio; depth images are not included.
- Original recordings are higher resolution (1280 x 720 and 848 x 530). This repository is the derived training dataset, not a full-resolution raw archive.
- Episodes retain their recorded start/end boundaries. No idle-time trimming was applied. Frames rejected by the existing synchronization checks or lacking a decodable image pair are omitted and listed in the provenance annotations.
Provenance and checks
annotations/episode_sources.json maps every exported episode to its original
session and episode, success mark, frame range and any omitted frames.
annotations/frame_sources.csv.gz maps each exported frame to its raw aligned-row
index and original wall-clock timestamp. Exported timestamps are regular 30 Hz
video timestamps; the original timestamps remain in this provenance table.
Some original capture intervals exceed one video frame. No synthetic frames or
interpolated robot states were inserted: the export plays the retained samples at
30 Hz, so these gaps are compressed in video time. Episodes with a source interval
above 100 ms are flagged by timing_gap_over_100ms; consult the source timestamps
when strict temporal fidelity is required.
The old collection tag block_stack_k0 was replaced only in this derived export
with the task instruction above. The original recordings were not modified.
All video streams were fully decoded and their frame counts checked. State/action
values were checked for finite values. For the start, middle and end of every
episode, loaded states/actions were compared against the original CSV and decoded
images were compared against the corresponding resized source images.
checksums.sha256 provides the SHA-256 of every published data/metadata/annotation
file and this dataset card.
Loading
from lerobot.datasets.lerobot_dataset import LeRobotDataset
successes = LeRobotDataset(
"Luckistt/nero-block-sort-stack-v1", revision="v3.0", video_backend="pyav"
)
failures = LeRobotDataset(
"Luckistt/nero-block-sort-stack-v1", revision="failures", video_backend="pyav"
)
No data license has been selected by the owner in this upload.