Skill-UMI

Handheld UMI-style gripper demonstrations with an Intel RealSense D455 wrist camera and two GelSight Mini visuo-tactile sensors (one per finger), contributed to Open-X-Tactile (OXT). Data accompanies the paper What to Attend, What to Keep: Skill-Conditioned Visuotactile Representation with Progress-Guided Event Memory (model: SkillFormer; project page).

Authors: Amir-Hossein Shahidzadeh, Seungjae Lee, Eadom Dessalene, Shanthosh Mageswari, Soroush Etemad, Furong Huang, Cornelia Fermüller, Yiannis Aloimonos (University of Maryland).

The zarr files follow the OXT / FTP-1 format. See the OXT website and the FTP-1 data processing guide for the full key definitions and coordinate conventions.

Sub-dataset	Episodes	Frames	Sub-task labels
bottle_cap	137	109,816	reach bottle cap, grasp bottle cap, reach bottle, align bottle cap, slide bottle cap, final twist, unlabeled
lamp_bulb	116	111,547	reach bulb, grasp bulb, reach lamp socket, align bulb, twist bulb, end
lentils_blind_search	144	100,751	reach box 1/2/3, grasp content 1/2/3, withdraw 1/2, reach puzzle box, align cube, insert cube

Each folder holds one zipped zarr (<name>.zarr.zip); unzip it to get <name>.zarr.

Keys (data/, all time-major, 30 Hz)

Key	Shape	Description
timestamps	(T,)	seconds since episode start
right_wrist_camera_rgb	(T, 224, 224, 3) uint8	wrist Intel RealSense D455 RGB
right_wrist_pose	(T, 6)	gripper pose (x, y, z, rotvec) in the SLAM world frame, OXT UMI gripper convention
right_hand_joints	(T, 1)	gripper width in meters
right_hand_joints_idx	(T, 1)	FAAS slot id, 28 (gripper)
right_tactile_data_gripper	(T, 2, 240, 320, 3) uint8	GelSight Mini images, one per finger, reference-subtracted (see below)
right_tactile_area_gripper	(T, 2)	tactile function-area ids [0, 1]
right_tactile_sensor_gripper	(T,)	GelSightMini
right_tactile_type_gripper	(T,)	image
sub_task_instruction	(T,)	sub-task label of each frame

meta/episode_ends gives the exclusive end index of each episode.

Tactile images are reference-subtracted

right_tactile_data_gripper holds difference images, not raw GelSight Mini frames:

stored = (frame - reference + 128) mod 256      # per pixel and channel, computed in uint8

• reference is the first frame of the raw GelSight video of each demonstration, captured before the episode's first frame while the gel is not in contact. Each finger has its own reference.
• A pixel with no change from the reference is 128. Because the subtraction was done in uint8, changes larger than ±127 wrap around.
• After subtraction the image is resized/cropped to 240x320.

The raw frames and reference images are not included.

Known pose issues (SLAM tracking loss)

right_wrist_pose comes from ORB-SLAM3 on the wrist camera. When tracking was lost, the pose is wrong for a stretch of frames while the images, tactile and gripper width are still valid. pose_quality.csv lists every affected frame range (per-episode and global frame indices). It flags three kinds of issues:

Issue	Rule	bottle_cap	lamp_bulb	lentils_blind_search
frozen	pose exactly constant for ≥ 1 s (30 frames)	28 / 137	42 / 116	46 / 144
jump	position changes > 10 cm in one frame	0	16	13
diverged	position > 2 m from the episode's median position	0	0	6

Most frozen segments are at the start of an episode, before SLAM had initialized. Examples of mid-episode tracking loss are lentils_blind_search episode 130 (frozen for frames 570–1015, ~15 s, while the gripper moves from the lentil box to the puzzle box) and episode 57 (drifts to ~5.9 m away from frame 666 on).

If you use the poses, drop or mask these ranges. A later version of this dataset will remove them.