Tac-Roboprompt

Teleoperated tabletop bimanual manipulation on an i2rt dual-arm platform, contributed to Open-X-Tactile (OXT). Each gripper has one Xsense Photon Tactile Sensor (stored identifier XenseOG); cameras comprise one fixed scene view and two wrist views.

v0.3 dataset

Task	Episodes	Frames
bottle	11	10,477
cup	20	24,973
lego	50	106,044
plug	13	19,121
pyramid	29	76,581
Total	123	237,196

This release removes the specified static-right-tactile episodes from v0.2, embeds sub_task_instruction in every task Zarr, and uses camera_main_rgb for the fixed scene camera. Its associated frame-number and timestamp arrays likewise use the camera_main_ prefix.

One complete Zarr ZIP per task

Each task is one complete ZIP64 archive, split by bytes into ordered volumes because of its size. All language, camera, tactile, robot-state and episode-boundary arrays are in that task archive. No .instructions.zip overlay is required.

tac-roboprompt/
  bottle/bottle.zarr.zip.001 ...
  cup/cup.zarr.zip.001 ...
  lego/lego.zarr.zip.001 ...
  plug/plug.zarr.zip.001 ...
  pyramid/pyramid.zarr.zip.001 ...

Each task directory has archive_manifest.json, listing every volume's order, size and SHA256, plus the SHA256 of the complete ZIP. Download all listed volumes. Unlike earlier independent ZIP chunks, these volumes must be concatenated before extraction:

# Example: run in the directory containing all cup volumes.
cat cup.zarr.zip.[0-9][0-9][0-9] > cup.zarr.zip
unzip cup.zarr.zip -d extracted_v03

This creates extracted_v03/cup.zarr/. Check volume hashes and the reconstructed ZIP hash against the manifest. Extract into a fresh directory so excluded episode chunks from an older release cannot remain. Current main contains only the complete v0.3 task archives; prior versions remain in repository history.

Read and use the data

Runnable Python example: examples/read_zarr.py

python -m pip install 'zarr>=2.18,<4' numpy
python examples/read_zarr.py extracted_v03/cup.zarr --episode 0 --frame 0

The example opens the dataset in read-only mode, checks episode boundaries, reads one frame and its instruction, and demonstrates all five image streams and arm/gripper state fields without loading an entire episode. Optional --preview-dir previews writes PNGs and requires Pillow.

Zarr v2 layout:

• data/camera_main_rgb: fixed scene RGB, (T,360,640,3).
• data/{left,right}_wrist_camera_rgb: wrist RGB, (T,240,320,3).
• data/{left,right}_tactile_data_gripper: (T,1,700,400,3) uint8 RGB, paired with area/sensor/type arrays.
• data/{left,right}_arm_joints: (T,6); gripper joints (T,1) with slot index 28.
• data/sub_task_instruction: per-frame task instruction, already inside each Zarr.
• data/timestamps: recorded host epoch seconds; use these rather than assuming uninterrupted 30 Hz.
• meta/episode_ends: cumulative exclusive frame endpoints.

See the FTP-1 processing format for the standard. OXT organizer metadata, annotations and check reports are delivered separately from this data repository.