2 episodes · 30 fps · 4 cameras · 640×400 av1
This dataset was created using LeRobot.
Dataset Description
Two short egocentric bimanual manipulation clips, pouring and folding (the last 15 s of each take, 450 frames each at 30 fps, 900 frames total). They were recorded with a head-mounted DECXIN stereo camera (175° lens) and two hand-held UMI-style grippers; each gripper carries a DECXIN wrist camera with an IMU, an ArUco marker cube and finger markers. Converted to LeRobotDataset v3.0 (lerobot==0.6.1); loads with LeRobotDataset("starpilot-ai/mobile_umi_test_2").
Same schema as starpilot-ai/mobile_umi_test, so the two can be mixed. Gripper A = left hand, gripper B = right hand. The task string is the task name (pouring, folding).
Poses come from our own pipeline: stereo visual odometry for the head, stereo-triangulated ArUco cubes for the grippers (short gaps bridged with each gripper's wrist IMU, orientation fused gyro + markers) and finger-marker tracking in the wrist cameras for gripper opening. This is a test release of the format; expect rough edges.
Source data and timing
- Head: 4000x1200 stereo at 60 fps, exposure 7.3 ms. Its layout is [strip][x 160..2080 = stereo_left][x 2080..4000 = stereo_right]. stereo_left is physically the right-hand lens (Basalt cam1).
- Wrists: 2080x1200 mono at 60 fps, exposure 4.2 ms (left) and 5.0 ms (right).
- The three cameras free-run, with no trigger. Each camera's exposure midpoint is mapped onto the shared host clock with a robust affine fit of its arrival stamps (fit residual p50 1.4 ms, p95 3.2 ms). The head stream is used at 30 fps (every 2nd frame) as the master timeline: one dataset frame per head frame. Each head frame is paired with the nearest 60 fps wrist frame. The pairing offset is always within half a 60 fps period: |dt| <= 8.3 ms, p95 about 7.9 ms. The cameras run at slightly different rates, so the offset sweeps that whole range. Unpaired head frames: pouring 1 per wrist camera (the first frame), folding 0 (left) and 1 (right).
- The wrist IMU was assigned to each gripper from the data: wrist gyro vs marker-cube rotation correlation was 0.95 and 0.99 (pouring) and 0.93 and 0.80 (folding). This also confirms the time alignment.
timestamp = frame_index / 30.exposure_tis the head exposure time on the host clock, relative to the first frame.
Coordinate frames and units
- world: gravity-aligned, per episode, in metres. z up comes from gravity measured by the two wrist IMUs, which agree within 0.9 deg (pouring) and 2.1 deg (folding). x is the head camera's viewing direction at frame 0, flattened onto the horizontal. The origin is the head (stereo_L) camera's position at frame 0. Each episode has its own world.
- head: the current stereo_L camera frame from stereo VO (
stereo_vo.py, KLT + stereo + PnP). It uses the Basalt calibration168_taha_calib_20261004.json(kb4 = OpenCV fisheye, baseline 60.16 mm). The head IMU is decoded and kept in the source folders, but it is not fused. - Grippers: the pose of the gripper body, i.e. the centre of the 4 marker-cube faces. A single cube model (face-to-body transforms, fitted on
pouring, where all 8 faces are seen) is used for both episodes, so the body frame is identical across episodes. Positions come from stereo-triangulated ArUco markers. Short gaps are bridged with the wrist IMU. Orientation is fused: the gyro carries fast motion and the markers carry the slow absolute part. - Smoothing: positions of the head and grippers get the lightest Gaussian smoothing (sigma 1 frame, +-2 frames).
- Poses are
x y z qx qy qz qw(scipy/ROS quaternion order, normalised to qw >= 0).
Markers (this gripper build)
All markers are DICT_4X4_50. Sizes were measured by stereo triangulation.
| gripper | hand | wrist camera | cube faces (49 mm) | finger markers (19.3 mm) |
|---|---|---|---|---|
| A | left | left_gripper (mono_1) | 2 3 4 5 | 11 (near-fixed finger), 10 (moving jaw) |
| B | right | right_gripper (mono_2) | 6 7 8 9 | 0 (near-fixed finger), 1 (moving jaw) |
Features
| key | shape | meaning |
|---|---|---|
| observation.images.stereo_left / stereo_right | 400x640x3 video | head stereo pair with the 160 px code strip removed (1920x1200 to 640x400) |
| observation.images.wrist_A / wrist_B | 400x640x3 video | wrist cameras: left gripper (A), right gripper (B) |
| observation.state | 23 | head pose in world (7), A pose in head frame (7), A opening_norm (1), B pose in head frame (7), B opening_norm (1) |
| observation.head.pose_world | 7 | head (stereo_L) pose in world |
| observation.gripper_{A,B}.pose_world | 7 | gripper pose in world |
| observation.gripper_{A,B}.pose_head | 7 | gripper pose in the current head frame |
| observation.gripper_{A,B}.opening | 3 | opening_norm (0..1, per episode between its 1st and 99th percentile), opening_mm, opening_deg |
| observation.imu_{A,B} | 6 | mean gx gy gz (rad/s) and ax ay az (m/s^2) of the wrist IMU over the frame interval. Raw IMU axes, no bias removal, gyro scale 28.65 LSB/(deg/s) |
| action | 20 | per gripper (A then B), absolute: the gripper pose at t+1 in the world frame (position, 3, m; rotation as 6D = first two rows of the matrix, 6) and its opening_norm at t+1 (1) |
| valid | 13 | 1 = measured, 0 = filled. Order: head_pose, gripper_A_pose, gripper_A_opening, gripper_B_pose, gripper_B_opening, img_stereo_left, img_stereo_right, img_wrist_A, img_wrist_B, imu_A, imu_B, action_A, action_B |
| pose_source | 2 | per gripper: 0 = missing, 1 = ArUco markers, 2 = IMU fill |
| head_vo_inliers | 1 | VO inlier count (a quality indicator) |
| exposure_t | 1 | seconds since the first exposure |
| timestamp, frame_index, episode_index, index, task_index | standard LeRobot columns |
valid, pose_source, head_vo_inliers and exposure_t are metadata, named so that LeRobot policies ignore them.
Missing-data convention
Every feature is dense. An unknown value holds the last valid value: forward fill, with a leading gap back-filled from the first valid value. Its valid entry is set to 0. An action is valid when the gripper pose is valid at t+1, so the last frame's action is always invalid. Mask losses with valid.
Coverage in this release
| pouring (450 frames) | folding (450 frames) | |
|---|---|---|
| head pose (stereo VO) | 450 | 450 |
| gripper A pose | 450 (markers 450) | 450 (markers 430, IMU fill 20) |
| gripper B pose | 450 (markers 450) | 450 (markers 436, IMU fill 14) |
| opening A / B | 450 / 450 | 434 / 449 |
| wrist images A / B | 450 / 450 | 450 / 449 |
On the full takes, IMU gap fill validated by hiding marker stretches gives a median error of 2.5–2.8 mm (pouring) and 8.6–9.1 mm (folding) for 0.5 s gaps.
Known limitations
Wrist cameras are not calibrated. The head unit's lens model stands in for both wrist cameras.
opening_mmandopening_degare therefore approximate: they are the marker-centre distance and the jaw angle, not the finger gap.opening_normis scale-free and is the value to use.Absolute scale rests on the Basalt stereo calibration. Cube markers triangulate at 49 mm and finger markers at 19.3 mm. The physical print sizes were not given.
The cameras are not hardware-synchronised. Host-clock pairing leaves up to ±8.3 ms between a head frame and its wrist frame. Wrist IMU samples are matched on each camera's own clock.
Homepage: [More Information Needed]
Paper: [More Information Needed]
License: apache-2.0
Dataset Structure
{
"codebase_version": "v3.0",
"fps": 30,
"features": {
"observation.images.stereo_left": {
"dtype": "video",
"shape": [
400,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 400,
"video.width": 640,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"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.images.stereo_right": {
"dtype": "video",
"shape": [
400,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 400,
"video.width": 640,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"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.images.wrist_A": {
"dtype": "video",
"shape": [
400,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 400,
"video.width": 640,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"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.images.wrist_B": {
"dtype": "video",
"shape": [
400,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 400,
"video.width": 640,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"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": [
23
],
"names": [
"head_x",
"head_y",
"head_z",
"head_qx",
"head_qy",
"head_qz",
"head_qw",
"A_x",
"A_y",
"A_z",
"A_qx",
"A_qy",
"A_qz",
"A_qw",
"A_opening",
"B_x",
"B_y",
"B_z",
"B_qx",
"B_qy",
"B_qz",
"B_qw",
"B_opening"
]
},
"observation.head.pose_world": {
"dtype": "float32",
"shape": [
7
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw"
]
},
"observation.gripper_A.pose_world": {
"dtype": "float32",
"shape": [
7
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw"
]
},
"observation.gripper_A.pose_head": {
"dtype": "float32",
"shape": [
7
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw"
]
},
"observation.gripper_A.opening": {
"dtype": "float32",
"shape": [
3
],
"names": [
"opening_norm",
"opening_mm",
"opening_deg"
]
},
"observation.imu_A": {
"dtype": "float32",
"shape": [
6
],
"names": [
"gx",
"gy",
"gz",
"ax",
"ay",
"az"
]
},
"observation.gripper_B.pose_world": {
"dtype": "float32",
"shape": [
7
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw"
]
},
"observation.gripper_B.pose_head": {
"dtype": "float32",
"shape": [
7
],
"names": [
"x",
"y",
"z",
"qx",
"qy",
"qz",
"qw"
]
},
"observation.gripper_B.opening": {
"dtype": "float32",
"shape": [
3
],
"names": [
"opening_norm",
"opening_mm",
"opening_deg"
]
},
"observation.imu_B": {
"dtype": "float32",
"shape": [
6
],
"names": [
"gx",
"gy",
"gz",
"ax",
"ay",
"az"
]
},
"action": {
"dtype": "float32",
"shape": [
20
],
"names": [
"A_x",
"A_y",
"A_z",
"A_rot6d_r00",
"A_rot6d_r01",
"A_rot6d_r02",
"A_rot6d_r10",
"A_rot6d_r11",
"A_rot6d_r12",
"A_opening_next",
"B_x",
"B_y",
"B_z",
"B_rot6d_r00",
"B_rot6d_r01",
"B_rot6d_r02",
"B_rot6d_r10",
"B_rot6d_r11",
"B_rot6d_r12",
"B_opening_next"
]
},
"valid": {
"dtype": "float32",
"shape": [
13
],
"names": [
"head_pose",
"gripper_A_pose",
"gripper_A_opening",
"gripper_B_pose",
"gripper_B_opening",
"img_stereo_left",
"img_stereo_right",
"img_wrist_A",
"img_wrist_B",
"imu_A",
"imu_B",
"action_A",
"action_B"
]
},
"pose_source": {
"dtype": "float32",
"shape": [
2
],
"names": [
"gripper_A",
"gripper_B"
]
},
"head_vo_inliers": {
"dtype": "float32",
"shape": [
1
],
"names": [
"inliers"
]
},
"exposure_t": {
"dtype": "float32",
"shape": [
1
],
"names": [
"seconds_since_episode_start"
]
},
"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": 2,
"total_frames": 900,
"total_tasks": 2,
"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": "vio_rig_handheld_bimanual",
"splits": {
"train": "0:2"
}
}
Citation
BibTeX:
[More Information Needed]
- Downloads last month
- 96