endotactile

Open-X-Tactile dataset contribution from Department of Electronic Engineering, The Chinese University of Hong Kong. Format and contribution instructions: OXT-Contribution and FTP-1 data processing.

Authors: Chi Kit Ng, Shuchang Liu, Siu Hing Lui, Xizi Wang, Ho Yin Chim, Yiu Cho Zhou, Jiewen Lai, Hongliang Ren.

Four semantic task families, eight experiment groups, 250 complete episodes, 98,402 aligned rows, approximately 20 Hz and 4,917.81 seconds of episode elapsed time. Fourteen standalone ZIP files contain Zarr v2 stores. Files stay below the FAT32 4 GiB limit; splitting is only at episode boundaries. UI presents the eight logical groups continuously. The upstream checker counts Zarr stores, so metadata.task is 14 rather than the semantic task count 4.

Protocol

• camera_ego_rgb: uint8 RGB [T,400,400,3], lossless pixels. Aligned chunks retain source Blosc/LZ4 bytes; shifted chunks are raw Zarr chunks inside lossless ZIP DEFLATE members.
• supplementary_joints: float32 [T,6], M1/M3/M4/M5/M6/M7 command-integrated estimates in controller-command-unit seconds. These are not measured joint angles, encoder readings or soft-body configuration.
• supplementary_action_commands: unchanged original target-speed controller commands, a custom auxiliary field. Future-state training targets are a different representation.
• supplementary_tactile_data_XJC620F: [T,2,1] normal force in kgf. Tensor channel 0: force_ch1=XJC CH1, image-left pink/red roller-support bracket. Tensor channel 1: force_ch2=XJC CH3, image-right orange roller-support bracket. Left/right refer to white_arrow_for_tactile_sensor.png. Channel locations were confirmed by the contributor on 2026-10-24; roles and units are also contributor-confirmed. Physical mounting locations do not assign canonical MTTS area IDs.
• supplementary_tactile_sensor_XJC620F and supplementary_tactile_type_XJC620F: per-row sensor/type strings.
• sub_task_instruction: per-row English instruction, generated from source phase/layout and contributor-confirmed base/grip goals.
• timestamps, meta/episode_ends and all original annotation sidecars are retained. Shard annotations use local rows plus source global ranges.
• provenance/source_arrays contains every original non-image array, including original state and invalid EM poses.

Contributor-authorized state normalization

Positive commands increase exported state. Base has factors [1,1,1,1,1,1]; all canpass layouts and cannotpass use [-1,1,1,1,1,1]; grip uses [1,1,1,1,-1,1]. Export=original*factor. No reintegration, offset shift, unit conversion or smoothing. Raw actions are unchanged. Inactive channels are retained without inferring their physical direction. Every shard includes factors and inverse-transform verification. Original state remains at provenance/source_arrays/colon_hand_joints.

Task meaning

Base: advance the endoscope through an unobstructed lumen. Grip: grasp an obstacle in front of the endoscope and bring it in front of the camera. Canpass/cannotpass retain source task/layout distinctions without asserting episode success. Eight episodes with original feedback remain included. The contributor confirmed completion of UI and VLM warning review on 2026-10-24; no exclusions were requested.

License

This dataset and its contributor-authored annotations/documentation are licensed under CC BY-NC 4.0. Attribute the authors listed above, link the license, indicate modifications, and use the material only for noncommercial purposes. See LICENSE.txt. The institutional logo, supplied reference material and third-party toolkit retain their applicable rights.

Force units were corrected from the previous erroneous N label to kgf on 2026-10-23. Values are unchanged raw GROSS readings. For SI use only, force_N = force_kgf * 9.80665; that conversion is not applied in this package.

Readiness

The contributor confirmed the submission scope on 2026-10-24: endoscopic RGB, six command-integrated motor-state estimates and two scalar normal-force channels in kgf. Original motor commands, language instructions, timestamps and episode annotations also remain included. All source EM poses are uninitialized. The contributor explicitly marked the 4D pose/calibration check NOT APPLICABLE for this scoped release on 2026-10-24. No valid pose, calibrated trajectory, performed 4D check, or independently verified organizer exemption is claimed. No vision-tactile image arrays are present. Row alignment is not proof of calibrated hardware synchronization.

Dataset repository: https://huggingface.co/datasets/kitRPAI/endotactile. License: CC BY-NC 4.0 (see LICENSE.txt). The contributor authorized publication and completed human review. OXT organizer acceptance has not yet been confirmed.

The physical sensor locations are complete: photo-left CH1 and photo-right CH3, both normal force in kgf. Official area values denote functional-area IDs, not arbitrary channel numbers. FTP-1 gives a two-pad gripper mapping example, but an explicit mapping for these endoscope feeding-roller supports was not found. The contribution guide permits three-field supplementary tactile data/sensor/type outside the current MTTS definition, while the pinned unmodified checker requires area and reports FAIL. Canonical area IDs were not invented. A separate local three-field validation passes all 14 shards (98,402 rows); this is not an official QC pass. See processing/supplementary_guideline_validation.json in the organizer packet.

The unchanged official VLM script completed actual local inference with unsloth/Qwen3-VL-8B-Instruct on rennc1 (RTX 5090, bfloat16): 14 stores, 70 montage pages, 1115 sampled rows. Dataset-level model statuses: {'warning': 14}. No cloud API was called; there were 0 API errors and 0 missing scores. These are original model screening results, not organizer acceptance. The contributor confirmed human review complete on 2026-10-24 and identified some false positives; model scores are preserved. See processing/official_qc_local_vlm_execution.json and the untouched oxt_check_output/vlm_review_result.json in the organizer packet. Contributor human review is complete, as recorded in contributor_review.json. No episode exclusions or individual warning decisions were supplied, so all data and original feedback remain. The area guide/checker discrepancy and contributor-declared 4D non-applicability are disclosed for organizer assessment.

Each dataset ZIP is independently extractable and contains a complete Zarr v2 store; no multipart joining is needed. Source and export hashes, transformations, raw official QC reports, model montages, source-code references and review records are included in the organizer package on the separate oxt-submission branch. That package is supplementary documentation; the 14 ZIPs in the eight dataset folders on main are the complete training data.