Datasets:
This dataset was created using LeRobot.
Dataset Structure
{
"codebase_version": "v3.0",
"fps": 20,
"features": {
"observation.images.front": {
"dtype": "video",
"shape": [
224,
224,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 224,
"video.width": 224,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.fps": 20,
"video.channels": 3,
"has_audio": false,
"video.g": 2,
"video.crf": 30,
"video.preset": 12,
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": {},
"is_depth_map": false
}
},
"observation.state": {
"dtype": "float32",
"shape": [
8
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw",
"gripper"
]
},
"action": {
"dtype": "float32",
"shape": [
8
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw",
"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
}
},
"total_episodes": 70,
"total_frames": 14024,
"total_tasks": 1,
"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": "fastumi_gripper",
"splits": {
"train": "0:70"
}
}
What the fields mean
| key | meaning |
|---|---|
observation.images.front |
GoPro frame, gripper masked out, ArUco tags inpainted away |
observation.state |
x, y, z, qx, qy, qz, qw, gripper — TCP pose in the table-tag frame, gripper width in metres (0 = closed, max 0.079) |
action |
same layout; action[t] = state[t+1], the last frame of an episode repeats itself |
In UMI, state and action come from the same arrays — sampler.py builds
action by concatenating [eef_pos, eef_rot_axis_angle, gripper_width], the
arrays the observation is also read from. Only the time direction differs: the
observation looks back, the action looks forward. Rotation was converted
axis-angle → quaternion here to match the other fastumi-* datasets.
Provenance
Egg-cracking demonstrations, FastUMI handheld gripper + GoPro HERO9 with a
third-party fisheye lens. Session 20260728: 86 raw demos, 71 with usable
SLAM trajectories, processed by the UMI SLAM pipeline (ORB-SLAM3 map → ArUco
tags → dataset plan → replay buffer) with --tcp_offset 0.131, the FastUMI
handle (lens-to-tip 0.14565 minus the screw-to-lens constant).
One episode was dropped before this export: 701 frames from demo
...16.45.51.140850, whose SLAM trajectory covers 3.55 m of path length
across a 0.26 m span (median step 1.4 mm, p99 104 mm).
Frame rate. The source holds one row per GoPro frame at 59.94 fps. Every
3rd row is kept here, which is exactly the policy's own action spacing
(obs_down_sample_steps: 3 → 50.05 ms) and lands at 19.98 Hz, declared as 20
because LeRobot's fps is an integer. Timestamps are i / 20, so the episode
clock runs 0.1% fast against the true GoPro clock.
Known caveats
- Residual SLAM transients. Scanning all 71 trajectories found 157 transient pose excursions (>8 mm off the smoothed path, returning within 4 frames); 85% sit at episode phase 0.5–0.7, which the IMU accelerometer independently places at the egg strike (impact peaks 6.6 g median). ORB-SLAM3 reports 100% tracked on every one of those frames, so nothing upstream filtered them. Typical magnitude 15–21 mm. Only the single worst episode was removed.
- Labels run ~2.7% short of metric. The session's map scale reads ~0.973 after correcting the +4.3% tag-PnP ranging bias measured on this rig.
Raw videos are not included.
- Downloads last month
- 103