Alan0928/franka_ft_oxt

Franka FT: four teleoperation tasks

A local, schema-compatible OXT-layout export of 125 curated successful demonstrations. It preserves the recorded robot and FT conventions; canonical wrist/TCP geometry and post-installation camera calibration have not been verified. This is not a claim of canonical training readiness or final OXT acceptance.

Public distribution prepared for Alan0928/franka_ft_oxt. The original local packaging records are preserved as provenance. No dataset license, institution, contact, or repository URL has been inferred.

Format references: Open-X-Tactile and the pinned FTP-1 official data format. Local OXT-QC checkout used for the metadata checks: cadbf5e7d1cfcec861a6d58779fc6d6e057f3126. These references do not imply formal acceptance.

Download and extract

The standard Zarr data is distributed in four ZIP archives under archives/. The duplicate expanded zarr/ tree is omitted from the repository because its 105,685 small files exceed convenient Hub layout limits. Every original Zarr file is preserved inside those ZIPs. All 500 original H5/SVO files, conversion tools, metadata, quality reports and previews are also included.

After downloading this repository into franka_ft_oxt/, run from that directory:

sha256sum -c SHA256SUMS
mkdir -p zarr
for archive in archives/*/*.zarr.zip; do
    unzip -n "$archive" -d zarr
done

The reading example below works after extraction. Each ZIP contains a single <task>.zarr/ directory, with identical bytes to the locally validated store. Allow approximately 27 GB of additional disk space for the extracted Zarr data.

publication/source_package_README.md and publication/source_package_SHA256SUMS preserve the original local README and complete-package checksums. The source package's upload_performed=false and local filesystem paths describe its pre-publication provenance, not the current availability of the repository. Publication changes only this README and the distribution checksum manifest; numerical data, images and raw files are unchanged. upload_manifest.json describes this distribution and its source package.

Contents and counts

Task	Episodes	Original controls	Standard rows	Full-rate FT samples
wipe board	38	14979	14941	105295
push button	27	4844	4817	34078
insert pen	30	11845	11815	81843
hang tape	30	10652	10622	74458
Total	125	42320	42195	295674

insert pen is the user's confirmed name for the original insert demonstrations. The prior user-excluded hang-tape episode at 2026-10-16 20:10:14 is not included; the replacement at 21:19:44 is included. Original task labels and selection evidence remain in metadata/selected_sources_audit.json.

• zarr/<task>.zarr/: Zarr format 2, data/* arrays and meta/episode_ends.
• native/<task>/<episode>/: complete original trajectory H5, full-rate FT H5, and original stereo SVO files.
• metadata/episodes/: source attributes, exact camera alignment, decoder calibration, and field mapping.
• metadata/array_schema.json: final task-wide array shapes, dtypes, chunks, compression, and field descriptions; metadata/field_mapping.json is explicitly a first-episode example.
• qc/: local numeric/preservation checks, OXT metadata prechecks, and any VLM dry-run artifacts.
• qc/previews/, examples/, and metadata/preview_metadata.json, when present: small human-viewing previews, separate from the original standard arrays; display resampling is documented in preview metadata.
• tools/ and requirements_oxt_export.txt: exact export, mapping, decoding, validation and packaging code plus offline regression tests.
• checksums/native_sha256.json: all native file hashes verified against the export manifest.
• SHA256SUMS: the uploaded distribution files, excluding this checksum file itself.
• archives/checksums.txt: SHA256 for each task archive.
• archives/<task>/<task>.zarr.zip: ZIP_STORED/ZIP64 archives of the already-compressed Zarr files; member sizes and CRC verified.

Native files are not duplicated in ZIP archives. Extract each task ZIP into a directory to obtain <task>.zarr/; its contents match the corresponding zarr/ store.

Array and timing conventions

right_arm_joints contains seven measured joint angles in radians. right_wrist_pose contains native robot-model flange position (local FR3 configuration: panda_link8) as xyz metres plus rotation-vector radians. Euler xyz is converted without changing the physical reference frame. right_hand_joints is normalized gripper closure, 0=open and 1=closed, with index 28; no physical-width scale is invented.

String arrays use explicit fixed-width Unicode with no variable-length UTF-8 filter in Zarr v2 storage. The exporter requests an empty filter list; the included fix_oxt_zarr_unicode.py repairs completed stores made with the older implicit filter behavior. Numeric and image arrays are not changed by that repair. Reader compatibility verification is separate from physical calibration.

True recorded commands are separate: action_joint_position, action_eef_pose, action_eef_pose_with_ft_offset, and command_dispatched. The EEF command is the requested target before IK, not FK of commanded joints and not an action inferred from a future observed state. Original action velocity fields are normalized per-step commands, not SI velocities. All 103 original row datasets are also retained under native__ names with / replaced by __.

The additional +63 mm pose fields shift the original EEF point along its local +Z with unchanged orientation. This is an installation length increment, not a complete TCP definition, sensor transform, or verified OXT canonical wrist transform.

timestamps is seconds since the original episode's first host control step; timestamp_unix_ns preserves exact host Unix nanoseconds. source_control_row maps each standard row back into the full original trajectory. Camera alignment requires SDK IMAGE nanoseconds floored to milliseconds to equal each recorded H5 camera timestamp in both cameras. Only the intersection is retained; no nearest-time guess, repeated final frame or fabricated observation is used. All original control rows remain in native H5, including those omitted from standard arrays.

FT snapshots use the original latest persisted sample before robot observation; their indices and host receive/snapshot times are retained. Full FT streams remain in each ft_sensor.h5. Robot-server seconds/nanoseconds are also preserved, but its clock was about 59.4–62.3 seconds behind the collection host and no offset correction is applied. Local FT/control comparisons use the common host monotonic clock; hardware synchronization is not claimed.

There are 3509 original rows with command_dispatched=false; those rows store hypothetical hold commands. Selected rows retain this flag. A true flag records dispatch, not confirmed asynchronous robot acceptance. The independent validator counted 3418 such rows in the standard arrays.

Force/torque and images

The sensor is SRI M3813A1 with M8228S1 interface. ft_wrench and supplementary_tactile contain [Fx,Fy,Fz,Mx,My,Mz] in the native sensor frame. N/Nm units are assumed from the device's stored calibration. The standard tactile four-piece set right_tactile_*_wrist_torque contains only [Mx,My,Mz], shape (T,1,3), area 15, sensor SRIM3813A1, type state. supplementary_tactile_note.json describes the full six-axis mapping.

No software bias subtraction, gravity compensation, or sensor-to-TCP/base transform is applied. Sensor mounting axes/origin, mass/COM and full transforms were not supplied. The device's calibration matrix and query responses remain in native metadata. Acquisition used host AT+GOD polling, approximately 70–99 Hz across these episodes; the device's reported internal 1000 Hz setting is not the saved receive rate. Packet-counter discontinuities are diagnostics, not a count of lost internal samples.

camera_main_rgb (external 36087771) and right_wrist_camera_rgb (16478870) are 640×360 RGB uint8, decoded using SDK LEFT rectified view. PNG staging and Blosc Zstd storage preserve the decoded, resized pixels losslessly; resizing itself is not reversible. Original full-resolution stereo recordings remain in SVO. Output intrinsics come from the decoder SDK calibration at the output resolution. Original H5 intrinsics and historical hand-eye/external extrinsics are preserved separately. The two active cameras' recorded calibrations date to December 2025; validity after FT installation is unconfirmed, and +63 mm was not silently applied to them.

Validation status

Independent preservation/numeric validation passed: qc/independent_validation.json. Scope: all chunks read.

OXT metadata task	Episodes	Rows	Hard errors	Warnings
hang_tape	30	10622	0	0
insert_pen	30	11815	0	0
push_button	27	4817	0	0
wipe_board	38	14941	0	0

Metadata checks establish layout consistency, not physical calibration or final OXT acceptance. The dataset-level compact metadata report still has overall_status=warning: it requests confirmation of the inferred sensor category and completion of human-authored publication/contact/reference fields. Those prompts are separate from the four passing task-level data checks above. VLM artifacts are local dry-run/montage evidence only; this package does not claim a VLM pass or a completed external model review.

• qc/vlm_dry_run/vlm_review.json: dry-run marker present.

No 4D visual review was run by this packaging workflow. Integrity checks cover stored bytes, not calibration accuracy.

Read one episode and its full-rate FT stream

from pathlib import Path
import h5py
import zarr

root = Path('franka_ft_oxt')
task = zarr.open_group(str(root / 'zarr/wipe_board.zarr'), mode='r')
data, meta = task['data'], task['meta']
end = int(meta['episode_ends'][0])  # first episode spans [0, end)
joints = data['right_arm_joints'][:end]
commands = data['action_joint_position'][:end]
rgb = data['right_wrist_camera_rgb'][0]
relative_native = str(meta['source_episode_paths'][0])
with h5py.File(root / relative_native / 'ft_sensor.h5', 'r') as ft:
    all_wrench = ft['samples/wrench'][:]  # all high-rate samples, (F, 6)
    receive_ns = ft['samples/receive_monotonic_ns'][:]
    indices = data['native__observation__ft_sensor__sample_index'][:end]
    recorded_snapshots = all_wrench[indices]

Install the pinned direct dependencies with python -m pip install -r requirements_oxt_export.txt. Reading Zarr/H5 requires no live hardware or ZED SDK. Re-decoding original SVO files requires a compatible ZED SDK/pyzed environment; SDK versions are recorded per episode.

Generating example MP4 previews also requires a system ffmpeg executable with the libx264 encoder on PATH; it is not installed by the Python requirements. Running the official OXT prechecks requires a separate OXT-QC checkout at the repository and commit recorded in metadata/format_reference.json.

Verify all delivered files from the package directory with sha256sum -c SHA256SUMS. Re-run offline mapping/decoder regressions with python -m unittest discover -s tools -p 'test_*.py'.

The Unicode compatibility repair includes its own synthetic tests: python tools/fix_oxt_zarr_unicode.py --self-test --zarr2-python /path/to/zarr2/python. Use a separate Zarr 2 reader environment for that cross-version check; the ordinary export/packaging environment's pinned Zarr version is listed in requirements.