starpilot-ai/mobile_umi_test_3

This dataset was created using LeRobot.

Dataset Description

One take of egocentric bimanual manipulation, split into three episodes:

episode	task	frames	duration
0	calibration (rig handheld, moved around the room before it goes on the head)	1200	40.0 s
1	kitting	2306	76.9 s
2	boxing	593	19.8 s

4099 frames at 30 fps. 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 a marker on each finger. Converted to LeRobotDataset v3.0 (lerobot==0.6.1); loads with LeRobotDataset("starpilot-ai/mobile_umi_test_3").

Same schema as starpilot-ai/mobile_umi_test and starpilot-ai/mobile_umi_test_2, so they can be mixed. The task string is the task name. This is a test release of the format; expect rough edges.

Privacy: bystanders' faces and computer screens (monitors, laptops) are lightly blurred in all four video streams (YOLO pose face keypoints and YOLO object detection, see meta/blur_report.json; in episode 2, face detections on the grippers' own fingers are dropped, and in wrist_A screens are not blurred 7.8-10.0 s in, where the detector landed on a gripper cube). The pose and opening labels were computed from the unblurred source.

Capture and timing

• Head: 4000x1200 stereo (two 1920x1200 views plus a 160 px timestamp strip), requested at 60 fps and delivered at a steady 30 fps; it is the trigger master.
• Wrists: 2080x1200 mono, hardware-triggered by the head camera's strobe, so every wrist frame is exposed together with its head frame (one dataset frame per head frame).
• Every frame carries the camera's own exposure start/end in its code strip; the wrist IMUs (300 Hz) come from the same strips.
• timestamp = frame_index / 30. exposure_t is the head exposure time relative to the episode's first frame.

Coordinate frames and units

• world: gravity-aligned, per episode, metres. z up from the gravity measured by the two wrist IMUs (they agree within 0.7°; the head camera's own IMU agrees with them within 0.7°). x = the head camera's viewing direction at the episode's first frame, flattened onto the horizontal. Origin = the head (stereo_L) camera at the episode's first frame.
• head: the stereo_L camera pose from stereo visual odometry. While worn, the head camera looks almost straight down at the workspace.
• grippers: the pose of the gripper body (centre of its marker cube), from stereo-triangulated 70 mm ArUco faces; short gaps are bridged with the wrist IMU and the orientation is fused gyro + markers.
• Poses are x y z qx qy qz qw (scipy/ROS order, qw >= 0).

Gripper opening

Both fingers of each gripper carry a 24 mm ArUco marker (id 24), seen by that gripper's wrist camera.

• opening_mm: 3D distance between the two finger-marker centres (PnP of each marker with the wrist camera's intrinsics). opening_norm = opening_mm scaled to 0..1 between its 1st and 99th percentile over the take. opening_deg is not provided (0).
• Frames where the moving finger's marker is not detected are recovered by template tracking along the finger's line of travel in the image (validated on detected frames: median error 0.5–0.7 px).
• Outliers (> 3 MAD from the ±7-frame median) are dropped and the series is lightly smoothed (5-frame median + Gaussian, σ = 1.5 frames).
• Gripper A's jaw moves little in this take (about 5 mm), so its opening signal is weak; gripper B's grasps are clear.

Features

key	shape	meaning
observation.images.stereo_left / stereo_right	400x640x3 video	head stereo pair, code strip removed
observation.images.wrist_A / wrist_B	400x640x3 video	wrist cameras of grippers A and 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 pose in world
observation.gripper_{A,B}.pose_world / pose_head	7	gripper pose in world / in the current head frame
observation.gripper_{A,B}.opening	3	opening_norm, opening_mm, opening_deg
observation.imu_{A,B}	6	mean wrist gyro (rad/s) and accel (m/s^2) over the frame interval, raw IMU axes
action	20	per gripper (A then B): pose at t+1 in world (position 3 + 6D rotation 6) and opening_norm at t+1 (1)
valid	13	1 = measured, 0 = filled (names in meta/info.json)
pose_source	2	per gripper: 0 = missing, 1 = ArUco markers, 2 = IMU fill
head_vo_inliers, exposure_t	1	metadata

Every feature is dense: unknown values hold the last valid value and their valid entry is 0. Mask losses with valid.

Coverage

	calibration	kitting	boxing
head pose	100%	100%	100%
gripper A pose (markers)	52% (37%)	100% (100%)	100% (100%)
gripper B pose (markers)	22% (11%)	100% (100%)	100% (97%)
gripper A / B opening	100% / 84%	98% / 85%	97% / 94%

In the calibration episode the grippers are mostly out of the head camera's view.

Known limitations

• The wrist cameras' intrinsics come from an earlier calibration of the same cameras, so opening_mm is approximate; opening_norm is the value to use.
• Head pose is visual odometry only (no loop closure).