CafeTac

CafeTac contains four robotic manipulation tasks, 110 complete trial episodes, and 120,321 joint-reference timesteps. Each episode includes seven arm joint states, one gripper joint state, head and wrist RGB images, and two Xense tactile pads with RGB, depth, and three-component force-grid observations.

The data is exported to the FTP-1-compatible Zarr structure for Open-X-Tactile. See the FTP-1 data-processing specification and ZARR_STRUCTURE.md for field definitions.

Release status

Dataset name: CafeTac. Format revision: 1.1.0. Public repository: https://huggingface.co/datasets/cheng000/CafeTac. Contact: pancheng000@gmail.com. The complete dataset author list and author homepages are provided below and in author_profiles.json. There is no dedicated CafeTac dataset paper. The teleoperation paper cited below is a robot-setup reference only. The contributor has not yet specified a data license. Visual-QC Warnings remain pending human review.

Files

README.md
metadata.json
author_profiles.json
ZARR_STRUCTURE.md
dataset_mapping.json
manifest.json
task_instructions.json
CafeTac/
  task1.zarr.zip
  task2.zarr.zip
  task3.zarr.zip
  task4.zarr.zip

Each ZIP contains its complete taskN.zarr/ directory. Extract all task archives under a common dataset directory. Sizes and SHA-256 checksums are in manifest.json.

Tasks and episodes

Task	Episodes	Timesteps	Instruction
task1	48	81,305	Pull the hard drive out of the server.
task2	8	6,864	Trials 1–4: take the bulb out of the tape roll and place it on the table. Trials 5–8: place it into another tape roll.
task3	26	14,432	Pick up one marker pen and place it in the clear plastic box.
task4	28	17,720	Move the pipe clamp from the left table to the right table.
Total	110	120,321

An episode spans one original trial from beginning to end, without trimming. Original trial numbers may have gaps. Episode indices within each task start at zero and must not be inferred by subtracting one from the original trial number.

Robot and camera mapping

• right_arm_joints (T,7): source Data1–7, in radians. Data1 is base left/right rotation; Data2–7 are the arm joints ordered from base to wrist.
• right_hand_joints (T,1): source Data8, the raw gripper angle in radians. right_hand_joints_idx is 28. Each trial starts with the gripper open.
• camera_ego_rgb (T,224,224,3): top/head camera, from arm4_gripper_usb_camera_stream_compressed.
• right_wrist_camera_rgb (T,224,224,3): wrist camera, from arm4_top_usb_camera_stream_compressed.

The right_* namespace represents one exported arm and gripper. It does not mean that both tactile sensors are on the right finger. The camera key alone does not establish ego/UMI classification or calibrated world-frame poses.

Tactile mapping

Both pads belong to the same gripper. For all three tactile groups:

Channel	Finger	Sensor serial	Functional area
0	Left	OG000470	0
1	Right	OG000144	1

• right_tactile_data_gripper (T,2,224,224,3): uint8 RGB images.
• right_tactile_data_depth (T,2,175,100): signed int16 raw counts.
• right_tactile_data_force_grid (T,2,35,20,3): signed int16 raw counts.
• Numeric validity masks, sensor names, modality types and area indices are retained. Representation aliases are XenseRGB, XenseDepth, and XenseGrid; an exact product model is not asserted.

Force units and component ordering/frame are unverified. Do not interpret raw force counts as Newtons. Display heatmaps are scaled previews, not the stored numerical data and not evidence of physical force saturation.

Timing and episode boundaries

Joint timestamps provide the reference timeline. Other streams are matched by nearest MCAP log time within the same trial. The previously verified maximum absolute matching offset is 28.9383 ms. This does not establish hardware capture synchronization or independent new sensor acquisitions at every timestep.

meta/episode_ends contains exclusive cumulative episode ends. For episode e, use [0, episode_ends[0]) if e=0, otherwise [episode_ends[e-1], episode_ends[e]). Original trial identifiers and source provenance are stored in meta/; matched stream indices/timestamps/offsets are stored in auxiliary/.

Quality-check status

• All four Zarr stores passed the OXT metadata/structure checks, with no dataset-level structural errors or warnings. Human-authored publication metadata remains incomplete.
• Direct Codex visual review covered all 110 episodes and 330 prepared montage pages: 41 Pass, 69 Warning, 0 Fail. No external VLM API was called. These are sampled visual assessments, not calibrated probabilities or full-frame validation. Organizer acceptance of this review method remains to be confirmed.
• The contributor confirmed successful bulb placement in task2/trial_6 and task2/trial_7. Earlier incorrect task-failure findings were withdrawn.
• Additional source-joint reviews found constant-value intervals with visible wrist/finger motion or ambiguous boundary motion. Source CSV values match the export; the cause remains under review. Separate joint concerns also exist for task1/trial_12 and trial_39, whose main visual status remains Pass.
• Real-coordinate 4D validation is not complete. Verified wrist/camera poses and kinematic calibration are not available; joint curves do not substitute for that validation. Ego/UMI applicability remains unresolved.

The organizer metadata/example package is prepared separately from these full data archives.

Tactile representation names and types

Data key	Sensor key value	Type	Per-pad representation
right_tactile_data_gripper	XenseRGB	image	RGB, 224 x 224 x 3, uint8
right_tactile_data_depth	XenseDepth	matrix	Depth, 175 x 100, int16 raw counts
right_tactile_data_force_grid	XenseGrid	matrix	Force grid, 35 x 20 x 3, int16 raw counts

All three groups use the same two physical pads, ordered left OG000470 (channel/area 0), then right OG000144 (channel/area 1). The names distinguish representations, not different hardware models. For OXT metadata, matrix representations are listed under sensor_array.

Archive	Size (GiB)	SHA-256
task1	31.320	d62e3189f3350980d387229090f2e6eb375401f08f44d6692d87f24cbfc4d9ad
task2	2.633	23be3320addd7142b20757d50ac86e59f0866ea0ff82d6b9bfb45c482e82364e
task3	5.466	8e053d88b96ec71b1b3d18845d785872fec5ba136aee11436f34109ae5d12769
task4	6.749	b9f43b621461c331f5f9d26954d7ecdbf26f6e6695b88651a04bae0f589318e0

Dataset authors and contact

• Cheng Pan
• Hung Hon Cheng (Paul)
• Yanyuan Qiao
• Dacheng Zhou
• Josie Hughes

Contact: pancheng000@gmail.com.

Robot-setup reference

Online Imitation Learning for Manipulation via Decaying Relative Correction through Teleoperation, Cheng Pan, Hung Hon Cheng, and Josie Hughes, 2025.

This paper is a reference for the robot setup; it is not a dedicated CafeTac dataset publication.