TacDexter

Real-world dexterous manipulation with fingertip tactile arrays, contributed to the Open-X-Tactile (OXT) dataset by the Department of Computer Science, University College London. The data follow the OXT / FTP-1 Zarr format; see the OXT website for the format specification.

A left Robotera XHand1 (12 actuated joints) is mounted on a handheld rig together with an Intel RealSense D405 wrist camera and a Pico VR controller. The operator moves the rig by hand (UMI-style) and drives the fingers with grasp poses blended in by the controller trigger (force-limited closing). Every fingertip carries the XHand1's built-in tactile array (10 x 12 taxels).

Contents

Sub-dataset (TacDexter/<task>.zarr.zip)	Trajectories	Frames	Minutes
stack_the_purple_cubes	44	26,206	14.6
put_all_cans_on_the_table_into_the_box	50	26,167	14.5
pick_up_the_watering_can_and_pour_into_the_red_bowl	45	16,198	9.0
insert_the_plug_into_the_socket	28	12,507	6.9
put_the_mustard_in_the_plate	49	11,145	6.2
pick_up_the_orange_and_put_it_into_the_bowl	50	8,732	4.9
Total (6 tasks)	266	100,955	56.1

Each ZIP holds one <task>.zarr/ directory (Zarr v2, ZIP_STORED / ZIP64). Verify with sha256sum -c SHA256SUMS, then unzip.

Keys (per task, T frames at 30 fps)

Key	Shape	Notes
data/timestamps	(T,) float64	seconds since the episode start
data/left_wrist_camera_rgb	(T, 224, 224, 3) uint8	D405 colour, short side scaled to 224 and centre-cropped
data/left_hand_joints	(T, 12) float32	measured joint angles (rad)
data/left_hand_joints_idx	(T, 12) int64	FAAS slots [1, 2, 3, 6, 7, 8, 12, 13, 17, 18, 22, 23]
data/left_tactile_data_fingertips	(T, 5, 10, 12) float32	normal force per taxel (0-255), thumb..pinky; see Known properties
data/left_tactile_area_fingertips	(T, 5) int64	MTTS areas [0, 1, 2, 3, 4] (fingertips)
data/left_tactile_sensor_fingertips	(T,) str	xhand1
data/left_tactile_type_fingertips	(T,) str	matrix
data/left_wrist_pose	(T, 6) float64	x, y, z (m) and rotation vector (rad), OXT frames
data/sub_task_instruction	(T,) str	the task instruction
meta/episode_ends	(E,) int64	exclusive episode ends

Streams are resampled onto the camera frames: joints are interpolated, tactile is the latest report at or before each frame, and wrist pose is the interpolated controller pose. The wrist orientation is the rig's calibrated wrist frame (z toward the fingers, x toward the thumb, y out of the palm) and the position is the controller's tracking point (no translation offset is calibrated), both relative to a level frame at the Pico headset standing fixed beside the workspace (OXT world: z forward, x right, y down). .zattrs["oxt_export"] records, for every episode, its source recording and time span.

Known properties

• Tactile resting levels. The hand does not zero its fingertip sensors: an untouched tip can read well above 0, and the level changes between sessions and with finger flexion.
  • In the three tasks recorded from the orange session on (pick_up_the_orange_and_put_it_into_the_bowl, put_all_cans_on_the_table_into_the_box, insert_the_plug_into_the_socket), where the middle tip rested at about 250-300 summed over its 120 taxels, each taxel's resting level (the 5th percentile over its episode) is subtracted and values within 2 of it are set to 0. The middle tip rests higher with the finger straight than bent, so in the cans and plug tasks an untouched, straight middle finger still reads up to about 220 summed (typically 100-135 at episode start).
  • The other three tasks are raw readings. An untouched tip reads 0 to about 100 summed (the pinky about 80), except in the stack_the_purple_cubes episodes recorded 2026-10-26 14:59-15:04, where the untouched index tip reads about 200-250 and, in the last three of them (from 2 s into 20261026_150237), the ring tip about 290.
  • .zattrs["oxt_export"]["tactile_rest"] states which applies to each file.
• Wrist tracking. When the Pico loses the controller it holds the position (the orientation keeps updating) and snaps back on re-acquisition; these gaps are bridged by interpolation, so positions inside a gap are estimates. Four smaller single-frame position steps of 5.6-6.8 cm remain (plug episode 6, orange episode 10, cans episodes 5 and 6).
• Joint ranges. Grasps are preset poses blended in by the controller trigger, so the thumb rotation (hand_joints_idx slot 1) stays within 1.47-1.62 rad and the index spread (slot 6) barely moves in many episodes.
• Excluded recordings. Four recordings were left out: one without camera frames and three in which the hand dropped off its EtherCAT bus mid-episode (joint readings frozen).
• Relabelled episodes. Five insert_the_plug_into_the_socket recordings (20261026_173559, _173619, _173652, _173712, _173801; .zattrs episode source) were recorded under the previous task's instruction and relabelled after checking their camera frames.

Quality checks

All six Zarr stores pass the official OXT-QC precheck_metadata.py with no errors or warnings. The VLM review (precheck_vlm.py, run with OpenAI gpt-4.1-mini) has no failing pages; its results are in the OXT submission package.

Authors and contact

Zhibin Li, Yiqin Li, Milo Carroll (Department of Computer Science, University College London). Contact: alex.li@ucl.ac.uk, yiqin.li.26@ucl.ac.uk

Citation

@misc{tacdexter2026,
  title        = {TacDexter: Teleoperated Dexterous Manipulation with Fingertip Tactile Arrays},
  author       = {Li, Zhibin and Li, Yiqin and Carroll, Milo},
  year         = {2026},
  howpublished = {Open-X-Tactile contribution, https://huggingface.co/datasets/RDLwicked/TacDexter}
}