Tacinf

Multimodal tactile manipulation: 45 tasks · 1,161 episodes · 157,770 frames. See Open-X-Tactile (OXT) for the data format and tools.

Subset	Collection	Episodes	Frames	Hz
UMI	Handheld UMI	175	23,332	20
RealPlus	Teleoperated specialized tasks	88	14,366	10
Sim / Replay	Simulation replay	178	26,590	20
Sim / ReplayPlus	Augmented simulation replay	720	93,482	20

Data

Tacinf/{UMI,RealPlus,Sim}/ contains one ZIP per task. Extract each ZIP beside itself to obtain its .zarr directory. Arrays use Zarr v2 + lossless Blosc/Zstd; tasks are described in English.

Field	Per-frame content
camera_main_rgb, right_wrist_camera_rgb	Main/wrist RGB, 224 × 224
right_wrist_pose	Current TCP position (m) + continuous rotation vector (rad)
right_hand_joints, right_hand_joints_idx	Gripper opening: 0 closed, 1 open; slot 28
right_tactile_data_image	Two current tactile RGB images, 224 × 224 each
right_tactile_data_force	Processed XYZ force for both pads, N
right_tactile_data_marker	220 absolute XY pixel coordinates per pad; not displacement or force
timestamps, sub_task_instruction	Episode-relative seconds and English instruction

right denotes the single arm/gripper. Tactile channels/area IDs 0, 1 retain source left/right pad order. Both pads use Xense G1 WS; Sim contains simulated signals. Force and marker groups use tactile type state. Motion targets are derived from pose/gripper trajectories; there is no separate 13D action array or force-action target.

Coordinates and processing

All poses use the robot Base reference, with local gripper axes Z forward, X down, Y left: R_export = R_LeRobot @ diag(-1,-1,1). Positions are unchanged by this axis mapping; rotation-vector continuity is preserved.

• RealPlus: measured link7 XYZ Euler feedback; mm → m, then p_TCP = p_link7 + R_link7 @ [0,0,0.232].
• Sim: the same link7-local +Z 0.232 m TCP offset is already included; it is not applied again.
• UMI: S_t = S_start @ inverse(U_0) @ U_t, followed by the same TCP conversion. S_start (mm, XYZ Euler degrees) is [251.542175, -5.547043, 207.964020, 82.510735, -88.606949, 97.333672]. This maps relative motion into a chosen robot reference, not a measured world registration.

Published LeRobot sampling and processed forces are retained: causal spike filtering for UMI/RealPlus, task820-cap15 for Replay, and frozen force alignment for ReplayPlus. No additional episode filtering; longer episodes remain. Camera extrinsics are unavailable; occasional source tracking/synchronization anomalies remain. Split augmented trajectories by common source motion to avoid train/test leakage.

Institution: The Hong Kong University of Science and Technology (Guangzhou)
Contributors: Qiwei Wu, Xiangyu Chen, Renjing Xu
Contact: debbiexinx@hkust-gz.edu.cn
License: CC BY-NC 4.0.