Source-linked dataset directory
Tactile datasets for robot learning
Find tactile and visuo-tactile datasets for robot learning. Filter by sensor, robot and task; compare primary download links, license status and split design.
Published 2026-07-20 | Updated 2026-09-20 by RoboSkin.ai Editorial Team
Structured dataset explorer
Filter tactile robotics datasets
Every row separates reported facts from unavailable fields. “Not stated” means the reviewed primary source did not provide enough evidence to fill that field.
Latest record review: 2026-09-19 / 20 records
Weekly brief / signup statusFree sample reportDiscuss a dataset research question
Showing 20 of 20 datasets
Select up to three records to compare. Your selection stays when filters change. Exports preserve access, license and source-review notes; they do not redistribute dataset files.
Scroll the table horizontally for all fields, or select records to read the comparison above.
| Dataset | Institution / year | Robot / sensor | Modalities / scale | Tasks / objects | Format / license | Primary links |
|---|---|---|---|---|---|---|
| Bench2Dex simulation demonstrationsUse the authors’ citation instructions on the primary paper or project page.Public file manifestAccess checked 2026-09-18Public, ungated provider manifest verified on 2026-09-18. Dataset payloads were not downloaded; completeness, checksums, schema validity and permission to reuse data remain unverified. Paper-scale demonstrations and hosted file counts are different units.Reviewed 2026-09-18 | Shanghai Jiao Tong University; Fudan University; The University of Hong Kong; Inspire Robots; Zhongguancun Academy; COWARobot Co. Ltd; Nanyang Technological University 2026 Original authors: Zhenjie Yang; Yideng Zhang; Dongjie Zhang; Chenyu Jiang; Xianshuai Liu; Yufeng Li; Zuhao Ge; Xingyu Jiao; Zheng Zhang; Kaiyu He; He Wang; Yuwen Zhong; Yi Deng; Muyun Jiang; Xianliang Huang; Haisheng Su; Donghang Zhang; Jian Zhang; Xue Yang; Hongyang Li; Zuxuan Wu; Yu-Gang Jiang; Xiaosong Jia; Junchi Yan First submitted 2026-09-14 arXiv v1 (2609.15726v1) Data origin: simulation | Robot: 12 simulated bimanual dexterous arm–hand embodiments; 26 task–embodiment settings Sensor: Surface-aligned simulated contact geometry; not measured output from a physical tactile sensor | Multi-view RGB; Joint state; Commanded actions; Object state; Surface-aligned tactile maps; Depth, 2D/3D boxes and occupancy when recorded Reported / previously documented scale: About 1,300 human-teleoperated simulation demonstrations in the v1 paper. Public hosting separately lists 5,200 HDF5 file paths; file paths are not a verified count of unique demonstrations. Public-file audit: 5,201 listed paths: 5,200 HDF5 files plus .gitattributes. Unique demonstrations, frames and bytes were not independently recounted. Independent check: Official project, v1 paper, repository root MIT license and provider manifests inspected. No payload download, checksum or HDF5 validation, simulation run or independent reproduction. | Bimanual tool use; Articulated-object interaction; Multi-stage manipulation; Controlled scene perturbations Objects: Task-specific tools, containers, articulated objects and assembly items; a deduplicated released object count is unknown. Generalization: Four scene-perturbation channels (None, Equi., Inv., Full). Multiple hand support is not zero-shot cross-hand transfer; no physical-sensor or sim-to-real validation claimed. | HDF5 episodes. The documented schema aligns frame_index, timestamp_ns and sim_step; action commands precede the next simulator step and observations follow it. Actions include names, control mode, validity and source. Supported fields are not guaranteed in every payload; README says depth is omitted in the current storage workflow. Dataset-file license: Unknown: no dataset card or dataset-file license found in the reviewed provider manifest. Code MIT is not a dataset license. Code / source terms: Repository root MIT applies to code; no separate dataset-file license or collection-wide model/asset license was verified. Code license: MIT at repository revision f96a8b2b4eb475483af66e9e03916b35bc43f1be. Asset license: Unknown collection-wide terms; upstream robot and object asset licenses require separate review. Model license: Unknown: no top-level weight license or model card verified; base-model terms may differ. Train / test split: Paper defines deterministic task-partitioned evaluation seeds and None / Equi. / Inv. / Full perturbation channels, with 50 rollouts per task/channel/policy. Released training/validation/test split membership and leakage remain unknown; a directory name is not a verified split. | |
| UniVTAC Encoder Pretraining CorpusUse the authors’ citation instructions on the primary paper or project page.Public file manifestAccess checked 2026-09-19Public contact-file listing verified on September 19, 2026. No standalone download establishing exact equivalence to the 205,826-sample paper corpus was verified; HDF5 payloads were not downloaded.Reviewed 2026-09-19 | ScaleLab, Shanghai Jiao Tong University; D-Robotics; ViTai Robotics; The University of Hong Kong; Nanjing University; Shenzhen University; Wuhan University; Fudan University; Tsinghua University 2026 | Robot: UniVTAC simulated gripper contact-generation setup; not a physical-robot trajectory corpus Sensor: Simulation-synthesized marker-based visuo-tactile observations | Raw tactile RGB image with markers; Marker-free tactile RGB image; Gelpad depth map; Projected 2D marker coordinates; Object pose as 3D translation plus quaternion Reported / previously documented scale: The paper reports 205,826 simulation samples across 14 geometric primitives. The September 19 provider inventory separately lists 638 contact-pretraining HDF5 files across 14 non-empty shape directories. File counts are not frame counts or proof of equivalence to the paper corpus; neither collection is the task benchmark, 400 paper policy-training trajectories or 450 physical demonstrations. Public-file audit: 638 listed contact HDF5 files across 14 non-empty shape directories. Equivalence to the paper's 205,826 contact samples is unverified; files are not frames. Independent check: Pinned provider file listing and code documentation inspected. No payload download, array validation, frame recount, training or hardware run. | Shape reconstruction; Contact-deformation prediction; Marker-position prediction; Relative object-pose regression; Tactile representation pretraining Objects: Fourteen geometric primitives ranging from convex forms such as spheres and cones to non-convex forms such as stars and cross-shapes, mounted on a standardized prism base for simulated contact generation. | The paper defines I_marked, I_pure, depth maps, projected 2D marker coordinates and a 7D object pose. The public contact/<shape>/hdf5/ paths were inventoried at revision e1aee7b0c95543b535e0146b2de3ee1bc6ddaabd; payload arrays, frame totals and splits were not validated. Dataset-file license: MIT in the provider dataset-card metadata; third-party assets and checkpoint terms require separate review. Code / source terms: The provider dataset card labels the repository MIT. This is separate from the code root Apache-2.0 license and does not resolve third-party asset or checkpoint terms. Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| UniVTAC Benchmark DatasetUse the authors’ citation instructions on the primary paper or project page.Public file manifestAccess checked 2026-09-19Public file listing verified on September 19, 2026. Match main/Isaac Sim 4.5 to isaac45, or isaac51/Isaac Sim 5.1 to isaac51. The official repository states task data are not cross-compatible and released policy checkpoints target 4.5. See the research guide for the downloadable 30-row inventory and payload-verification limits.Reviewed 2026-09-19 | ScaleLab, Shanghai Jiao Tong University; D-Robotics; ViTai Robotics; The University of Hong Kong; Nanjing University; Shenzhen University; Wuhan University; Fudan University; Tsinghua University 2026 | Robot: Simulated Franka Panda with a parallel-jaw gripper and bilateral simulated GelSight Mini sensors Sensor: Bilateral simulated GelSight Mini sensors through UniVTAC and TacEx | Head RGB image; Wrist RGB image; Robot end-effector and joint state; Robot action and episode metadata; Bilateral tactile RGB; Bilateral marker-overlaid tactile RGB; Bilateral tactile depth; Bilateral marker coordinates; Bilateral tactile-sensor pose Reported / previously documented scale: September 19 file inventory at revision e1aee7b0c95543b535e0146b2de3ee1bc6ddaabd lists 800 HDF5 paths under isaac45 and a separate 800 under isaac51, 100 per task. The historical August revision 172331dbbce95bc04c3e59b22f32dc72ba5561ae contained 800 HDF5 episodes, approximately 125.43 GB, with metadata marking 763 successful and 37 non-success episodes. Historical size/outcome counts were not revalidated for the current versions. Task files are distinct from encoder samples, policy-training trajectories and physical demonstrations. Public-file audit: 2,344 paths in the entire provider repository, shared with the contact and checkpoint collections. Task-only inventory: 800 HDF5 paths in isaac45 and 800 in isaac51, 100 per task in each version. Bytes, frames and unique demonstrations were not recounted. Independent check: Official hosting API and file manifest inspected at the pinned revision. No dataset payloads downloaded; no checksums, schema validation or training run. | Lift Bottle; Pull-out Key; Lift Can; Put Bottle in Shelf; Insert Hole; Insert HDMI; Insert Tube; Grasp Classify Objects: Task-specific bottles, cans, keys and locks, shelf-placement objects, peg-and-hole components, HDMI connectors and ports, tube fixtures, and tactile-classification shapes as defined by the eight benchmark tasks. | Versioned isaac45/<task>/hdf5/ and isaac51/<task>/hdf5/ directories with task metadata. Collection documentation describes visual observations, robot state and bilateral tactile RGB, marker, depth and pose fields. Current payload schemas were not validated. The August Dataset Viewer CastError is historical; current viewer status was not retested. Dataset-file license: mit in the provider dataset-card metadata at the pinned revision; individual asset terms were not inspected. Code / source terms: MIT as displayed by the current Hugging Face dataset card. The official GitHub repository root LICENSE is separately Apache-2.0; its README’s MIT statement conflicts with that root license and is not used as the repository-license authority. Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| T-Rex Tactile-Reactive Dexterous Manipulation DatasetUse the authors’ citation instructions on the primary paper or project page.Public file manifestAccess checked 2026-09-13The 5,464-episode, approximately 50-hour LeRobot v3.0 subset is publicly browsable and downloadable on Hugging Face. The official repository provides a dataset quickstart and states that the full 100-hour pretraining and midtraining corpus is not part of the current release.Reviewed 2026-08-22 | UC Berkeley; NVIDIA; Stanford University; Panasonic; Sapienza University of Rome; ItalAI 2026 | Robot: Fixed-base bimanual Dexmate Vega-1 with two 22-DoF Sharpa Wave dexterous hands Sensor: ZED X Mini head camera; Two ZED X One S wrist cameras; Ten fingertip image-based tactile sensors, five per hand | Head RGB video; Bilateral wrist RGB video; Robot state; Current and target joint positions; Raw fingertip grayscale tactile video; Estimated tactile deformation maps; Estimated 6D fingertip wrenches; Frame and episode metadata Reported / previously documented scale: The paper and official repository report a complete 100-hour collection with 7,700-plus trajectories, 22 motor primitives, and more than 200 objects. The public Hugging Face card specifies 5,464 episodes, 5,473,459 frames at 30 FPS (approximately 50 hours), 5,370 language-annotated trajectories, 22 motor primitives, and 207 objects. The public subset must not be described as the complete 100-hour corpus. Public-file audit: 7,903 paths in the provider file manifest, including metadata. Bytes, frames and episodes were not independently recounted. Independent check: Official hosting API and file manifest inspected at the pinned revision. No dataset payloads downloaded; no checksums, schema validation or training run. | Bimanual tactile-reactive motor primitives; Dexterous manipulation; Force-sensitive contact; Deformable-object interaction; Vision-language-action midtraining Objects: The current public dataset card reports 207 unique canonical objects and publishes composition metadata, but RoboSkin did not independently audit every object label or collapse them into an inferred category taxonomy. | LeRobotDataset v3.0 with head and bilateral wrist videos, robot state, current and target joint positions, ten fingertip tactile streams, deformation maps, 6D wrenches, and episode metadata. The official repository provides selective download, inspection, and 3D replay examples. Dataset-file license: mit in the provider dataset-card metadata at the pinned revision; individual asset terms were not inspected. Code / source terms: MIT according to the current Hugging Face dataset card. Repository code is also MIT, but upstream robot assets, dependencies, and any files outside the public package retain their own terms. Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| EgoTouchUse the authors’ citation instructions on the primary paper or project page.Public file manifestAccess checked 2026-09-13The official Hugging Face repository exposes scene folders and split metadata. The official GitHub repository says the upload is in progress and currently accessible files may be incomplete, so completeness and the exact version must be checked before reuse.Reviewed 2026-08-22 | Harbin Institute of Technology, Shenzhen; Meituan Academy of Robotics; Tsinghua Shenzhen International Graduate School, Tsinghua University 2026 | Robot: Human-wearable bimanual collection; no robot platform is used in the EgoTouch capture pipeline Sensor: Head-mounted wide-angle camera; Dual wrist cameras; Bimanual pressure-sensing gloves with dense palm pressure maps; Bimanual 3D hand-pose capture | Head-mounted egocentric RGB; Dual wrist-mounted RGB; Bimanual 3D hand pose with 42 joints; Continuous bilateral tactile pressure maps; Camera and wrist poses; Frame timestamps Reported / previously documented scale: 208 manipulation tasks across 1,891 episodes in indoor and outdoor environments. The official repository reports approximately 2 million frames and more than 1,000 objects; those repository statistics are author-reported and the repository still warns that the dataset upload may be incomplete. Public-file audit: 22,897 paths in the provider file manifest, including metadata. Bytes, frames and episodes were not independently recounted. Independent check: Official hosting API and file manifest inspected at the pinned revision. No dataset payloads downloaded; no checksums, schema validation or training run. | Bimanual hand-object manipulation; Egocentric vision-to-touch prediction; Contact prediction under visual occlusion; Multi-view tactile estimation Objects: More than 1,000 objects across home, office, outdoor, retail, and workbench scenes are reported by the official repository; a complete reviewed object taxonomy was not verified. | The official repository documents variable-length 30 FPS HDF5 episodes containing three 640 x 480 RGB views, bimanual hand-pose arrays, bilateral pressure grids, camera poses, masks when available, metadata, and timestamps. The Hugging Face repository also exposes scene folders and split.json. Dataset-file license: Unknown: a dataset-file license has not been verified separately from paper or code terms. Code / source terms: A separate dataset-file license was not verified. The official GitHub repository is MIT licensed, but that project-code license is not generalized here to the hosted EgoTouch data files. Train / test split: Split-named files listed: split.json. Contents and leakage were not validated. | |
| PRISM: Precision and contact-rich Real-world Industrial Skill dataset with Multimodal sensingUse the authors’ citation instructions on the primary paper or project page.Announced; release not verifiedAccess checked 2026-08-22The arXiv abstract says the dataset is open-sourced at the project page, but as of 2026-08-22 the official Dataset button is disabled and labeled “soon”; the GitHub repository contains website and paper assets but no dataset files, releases, or dataset license. Record PRISM as announced, download pending.Reviewed 2026-08-22 | State Key Laboratory of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University; Delta Intelligence; PKU-Wuhan Institute for Artificial Intelligence; Hubei Humanoid Robot Innovation Center Co., Ltd.; China Academy of Information and Communications Technology 2026 | Robot: Dual Franka Emika Panda tracker platform; Bimanual Realman RM75-6F exoskeleton platform; LEJU upper-body humanoid VR platform Sensor: Multi-view RGB-D cameras; 6DoF force/torque sensing; Unnamed visuotactile gripper and visuotactile dexterous hand on a subset of Franka episodes | Multi-view RGB; Depth; Visuotactile image when available; 6DoF force/torque when available; Robot joint angle and torque; End-effector Cartesian pose; Gripper state; Calibration and timestamps; Human control signal Reported / previously documented scale: 5,000+ robot trajectories paired with 5,000 human demonstrations; 45+ hours; 25+ industrial tasks; approximately 27M images across visual and visuotactile streams. The paper does not disclose the number or share of tactile-equipped episodes. Public-file audit: Unknown: no independently recounted public-file total in this record. Independent check: Existing source review only. No new dataset-file download, checksum validation, training run, or hardware test in this audit. | Electronic component plug and unplug; Caliper packaging; Conveyor sorting; NIST Assembly Task Board operations; Industrial installation and assembly; Long-horizon compositional procedures Objects: Industrial components, assembly hardware, packaging items, conveyor-sorted materials, and task-board objects; the source does not provide one complete released object inventory. | The paper describes timestamped episode files in a common cross-platform schema and reports conversion to LeRobot v3.0 for its experiments. The unreleased public package and final file inventory could not be inspected. Dataset-file license: Unknown: a dataset-file license has not been verified separately from paper or code terms. Code / source terms: Dataset-file license not published on the reviewed project page or GitHub repository; the article license does not license unreleased data files Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| SoftVTBenchUse the authors’ citation instructions on the primary paper or project page.Public file manifestAccess checked 2026-09-13The official repository links public code and the Hugging Face dataset mirror. The Hugging Face card reviewed on 2026-08-22 documents 4,000 hosted demonstrations, while the older GitHub README still lists 1,628 demonstrations and 33 assets. Pin the exact dataset revision because the first-party release documents currently disagree.Reviewed 2026-08-22 | Tuojing Intelligence; Tsinghua University; King's College London; Southeast University; Stevens Institute of Technology; Hong Kong University of Science and Technology (Guangzhou); University of Manchester; Simple AI; Imperial College London; Carnegie Mellon University; Zhejiang University; Beihang University; University of Hong Kong 2026 | Robot: Simulated Franka arm with Panda parallel-jaw gripper in Isaac Sim and Isaac Lab Sensor: Simulated bilateral GelSight Mini profiles rendered through TacEx, Taxim, and FOTS | Multi-view RGB; Dual-finger tactile RGB; Tactile marker motion; Proprioception; Language; Continuous and binary gripper actions; Evaluator-only FEM state Reported / previously documented scale: Latest paper: 4,000 expert demonstrations across 40 tasks and more than 50 assets. Hugging Face dataset-card revision fd2793a documents four subsets with 10 tasks and 100 successful demonstrations each, or 4,000 demonstrations total. The GitHub README last changed 2026-07-22 still lists an earlier 1,628-demonstration, 33-asset state, so the first-party release documents are not synchronized. Public-file audit: 16,060 paths in the provider file manifest, including metadata. Bytes, frames and episodes were not independently recounted. Independent check: Official hosting API and file manifest inspected at the pinned revision. No dataset payloads downloaded; no checksums, schema validation or training run. | Deformable-object grasp and placement; Rigid-twin matched control tasks; Object-variation generalization; Spatial-variation generalization; Vision-only and visuo-tactile policy comparison Objects: Ten volumetric deformable manipulated-object designs and matched rigid twins; the latest paper reports more than 50 assets across the complete benchmark inventory. | HDF5 trajectories with synchronized policy observations, robot actions, simulator state, and evaluator fields, plus third-person, wrist, and bilateral tactile MP4 videos documented by the current dataset card. Dataset-file license: apache-2.0 in the provider dataset-card metadata at the pinned revision; individual asset terms were not inspected. Code / source terms: Apache-2.0 as shown on the current Hugging Face dataset card; upstream simulator, robot, tactile-runtime, and third-party asset terms may differ Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| RoboTacDex: A Dexterous Visual-Tactile-Action Dataset for Humanoid ManipulationUse the authors’ citation instructions on the primary paper or project page.Announced; release not verifiedAccess checked 2026-08-22The June 30, 2026 v1 paper says the dataset will be open-sourced soon. No official project page, repository, downloadable package, or dataset-file license was verified on 2026-08-22; record access as announced, not released.Reviewed 2026-08-22 | Fudan University; ByteDance Intelligent Creation; Shanghai AI Laboratory; The Chinese University of Hong Kong 2026 | Robot: Unitree G1 humanoid with dual arms totaling 14 DoF and two BrainCo Revo2 Tactile dexterous hands totaling 12 hand DoF as counted by the paper; lower limbs and waist fixed during collection Sensor: Four RGB-D cameras, including two Intel RealSense D435i cameras for the head and third-person views; Two BrainCo Revo2 Tactile dexterous hands measuring fingertip normal force, tangential force and direction, and self-capacitance proximity | Four-view RGB at 640 x 480; Four-view depth at 640 x 480; Arm and finger joint state; Arm and finger joint action; Bilateral fingertip normal force; Bilateral fingertip tangential force and direction; Self-capacitance proximity; Natural-language semantic annotation Reported / previously documented scale: More than 6,000 physical-robot trajectories totaling approximately 25 hours, covering 19 tasks, an author-reported 23 skills, and 22 objects. Figure 4 exposes 22 discernible atomic-skill labels, so the 23-skill headline has an unresolved internal source discrepancy. Trajectories are recorded at 30 Hz; tactile and dexterous-hand joint-state messages are published at 100 Hz and recorded locally at 30 Hz with the other modalities. Public-file audit: Unknown: no independently recounted public-file total in this record. Independent check: Existing source review only. No new dataset-file download, checksum validation, training run, or hardware test in this audit. | Basic grasping, placing, and pushing; Articulated-object manipulation; Dual-arm collaborative manipulation; Fine manipulation; Humanoid-interactive manipulation; Pick and place a pear; Turn a page; Insert a book into a document bag; Unscrew a bottle cap Objects: Twenty-two objects spanning articulated, constrained, container, functional, and graspable categories, including kitchen food and office supplies as well as rigid and deformable objects. The paper does not publish a complete machine-readable object inventory. | The v1 paper describes 30 Hz synchronized trajectories containing four 640 x 480 RGB-D views, arm and finger states and actions, bilateral tactile signals, and semantic annotations. It does not state the public package format, directory schema, compression, checksums, or train-validation-test file layout. Dataset-file license: Unknown: a dataset-file license has not been verified separately from paper or code terms. Code / source terms: Dataset-file license not stated. The arXiv article license does not license unreleased dataset files. Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| HT-BenchUse the authors’ citation instructions on the primary paper or project page.Announced; release not verifiedAccess checked 2026-08-22The v2 preprint says the authors will release all data, evaluation protocols, pretrained weights, and training/testing scripts. No dedicated downloadable dataset package, official repository, or artifact license was verified on 2026-08-22; record access as announced, not released.Reviewed 2026-08-22 | Beihang University; Rimbot; ShanghaiTech University; Tsinghua University; Chinese Academy of Sciences; BUPT 2026 | Robot: Single reported dexterous full-hand collection and downstream-evaluation platform; hardware model not stated in the reviewed HTML Sensor: Single reported full-hand tactile sensing pipeline; hardware model not stated in the reviewed HTML | Egocentric RGB; Full-hand tactile pressure maps Reported / previously documented scale: 10M RGB frames; 7.8M tactile frames; 226 tasks Public-file audit: Unknown: no independently recounted public-file total in this record. Independent check: Existing source review only. No new dataset-file download, checksum validation, training run, or hardware test in this audit. | Tactile similarity retrieval; Masked tactile inpainting; Vision-to-tactile synthesis; Tactile frame prediction Objects: Home, electronics workbench, chemistry lab, retail, workbench, outdoor, and other scenes; object count not stated. | Synchronized egocentric RGB and full-hand tactile maps; the benchmark pipeline normalizes tactile maps to 224 x 224. Dataset-file license: Unknown: a dataset-file license has not been verified separately from paper or code terms. Code / source terms: Dataset, model-weight, and code licenses not stated; the paper is CC BY 4.0, which does not establish a license for unreleased artifact files Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| RCT: Robotic Contact TactileUse the authors’ citation instructions on the primary paper or project page.Download route linked; files not verifiedAccess checked 2026-08-16Dataset, split tools, and evaluation code are publicly linked by the official project page.Reviewed 2026-08-16 | TU Dresden; ScaDS.AI Dresden/Leipzig; LASR Lab 2026 | Robot: Robot arm with rotating three-sensor adapter Sensor: DIGIT vision-based tactile sensor | Tactile image; Material RGB image; Language descriptors; Normal force; Indentation depth Reported / previously documented scale: 29,279 tactile frames; 1,832 contact sequences; 122 materials; 3 DIGIT sensors Public-file audit: Unknown: no independently recounted public-file total in this record. Independent check: Existing source review only. No new dataset-file download, checksum validation, training run, or hardware test in this audit. | Material generalization; Touch-to-text retrieval; Touch-to-vision retrieval; Sensor-disjoint evaluation Objects: 122 industrial reference materials in 7 categories. | Ordered contact sequences with material, category, sensor, position, depth, force, image, and descriptor metadata. Dataset-file license: CC BY 4.0 dataset; Apache-2.0 applies to code, according to the existing project review. Code / source terms: CC BY 4.0 dataset; Apache-2.0 code Train / test split: Existing source review documents split tools. File-level membership and train/test leakage were not independently validated. | |
| TactiDexUse the authors’ citation instructions on the primary paper or project page.Access unknownAccess checked 2026-08-16The official project page documents the benchmark and demonstrations; no separate dataset download URL was verified.Reviewed 2026-08-16 | ShanghaiTech University; InstAdapt 2026 | Robot: Human demonstration capture; Bimanual Franka Inspire deployment Sensor: Whole-hand tactile glove | Whole-hand pressure map; Hand kinematics; Object 6D pose; Language description; Task-phase annotation Reported / previously documented scale: Not stated on the reviewed project page Public-file audit: Unknown: no independently recounted public-file total in this record. Independent check: Existing source review only. No new dataset-file download, checksum validation, training run, or hardware test in this audit. | Single-hand dexterous manipulation; Bimanual manipulation; Human-to-robot skill transfer Objects: Household and tool objects shown in project demonstrations; total object count not stated. | Synchronized tactile pressure, hand-object kinematics, 6D pose, language, and hierarchical task annotations. Dataset-file license: Unknown: a dataset-file license has not been verified separately from paper or code terms. Code / source terms: Dataset license not stated on the reviewed project page Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| FreeTacManUse the authors’ citation instructions on the primary paper or project page.Public file manifestAccess checked 2026-09-13The official project page links the dataset, code, hardware guide, and mirror.Reviewed 2026-08-16 | Shanghai Innovation Institute; The University of Hong Kong; Shanghai Jiao Tong University; Fudan University 2025 | Robot: Robot-free wearable gripper; Piper; Franka Sensor: Modular LED-based visuo-tactile sensor | Wrist RGB video; Visuo-tactile video; Tool-center-point pose; Gripper distance Reported / previously documented scale: More than 3M visuo-tactile image pairs; more than 10K trajectories; 50 tasks Public-file audit: 24,915 paths in the provider file manifest, including metadata. Bytes, frames and episodes were not independently recounted. Independent check: Official hosting API and file manifest inspected at the pinned revision. No dataset payloads downloaded; no checksums, schema validation or training run. | Contact-rich demonstrations; Imitation learning; Tactile pretraining; Visuo-tactile manipulation Objects: Fifty contact-rich task categories; examples include fragile handling, insertion, stamping, texture classification, and calligraphy. | MP4 wrist and tactile videos plus timestamped trajectory files containing TCP pose, quaternion, Euler angles, and gripper distance. Dataset-file license: mit in the provider dataset-card metadata at the pinned revision; individual asset terms were not inspected. Code / source terms: MIT License Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| Humanoid Visual-Tactile-Action DatasetUse the authors’ citation instructions on the primary paper or project page.Access unknownAccess checked 2026-08-16The paper describes the dataset; no official public download URL was verified on 2026-08-16.Reviewed 2026-08-16 | Gwangju Institute of Science and Technology 2025 | Robot: Humanoid teleoperation platform; Two Inspire RH56-DFX dexterous hands Sensor: 1,062 tactile sensors per hand; Piezoresistive tactile carpet | Egocentric RGB; Third-person RGB; Dense tactile pressure; Arm and finger proprioception; Robot action; External pressure heatmap Reported / previously documented scale: 101.9K synchronized samples; approximately 77-80 episodes per task Public-file audit: Unknown: no independently recounted public-file total in this record. Independent check: Existing source review only. No new dataset-file download, checksum validation, training run, or hardware test in this audit. | Towel strong pressure; Towel weak pressure; Sponge strong pressure; Sponge weak pressure Objects: Two deformable soft objects (towel and sponge), with rigid-object comparison data described in the paper. | Synchronized visual, 2,124-channel hand tactile, proprioceptive, action, and external pressure signals; public file format not stated. Dataset-file license: Unknown: a dataset-file license has not been verified separately from paper or code terms. Code / source terms: Dataset access terms not stated on the reviewed paper page Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| Sparsh-X Multisensory Touch ResourceUse the authors’ citation instructions on the primary paper or project page.Access unknownAccess checked 2026-08-16The paper documents the training resource; a dedicated public dataset URL was not verified.Reviewed 2026-08-16 | FAIR at Meta; University of Washington; Carnegie Mellon University 2025 | Robot: Robot manipulation platforms used for insertion and in-hand rotation Sensor: Digit 360 | Tactile image; Audio; Motion; Pressure Reported / previously documented scale: Approximately 1M unlabeled contact-rich interactions Public-file audit: Unknown: no independently recounted public-file total in this record. Independent check: Existing source review only. No new dataset-file download, checksum validation, training run, or hardware test in this audit. | Physical-property inference; Plug insertion; In-hand rotation; Tactile adaptation Objects: Diverse manipulation interactions; object count and category inventory not stated on the reviewed paper page. | Four synchronized Digit 360 modalities; public file format not stated on the reviewed paper page. Dataset-file license: Unknown: a dataset-file license has not been verified separately from paper or code terms. Code / source terms: Dataset license and download URL not stated on the reviewed paper page Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| Touch and GoUse the authors’ citation instructions on the primary paper or project page.Download route linked; files not verifiedAccess checked 2026-08-19The paper and project describe a human-collected in-the-wild visuo-tactile dataset; verify the current download terms before reuse.Reviewed 2026-08-19 | University of Michigan; Carnegie Mellon University 2022 | Robot: Human-operated GelSight collection device Sensor: GelSight | Egocentric RGB video; GelSight tactile video; Material labels Reported / previously documented scale: Approximately 13.9K detected touches; approximately 3,971 object instances; 20 material categories Public-file audit: Unknown: no independently recounted public-file total in this record. Independent check: Existing source review only. No new dataset-file download, checksum validation, training run, or hardware test in this audit. | Visuo-tactile feature learning; Material recognition; Tactile-driven image stylization; Future tactile prediction Objects: Rigid and deformable objects across indoor and outdoor environments; the paper reports 20 material categories. | Synchronized egocentric video and GelSight tactile recordings with detected-touch timing and material annotations. Dataset-file license: Unknown: a dataset-file license has not been verified separately from paper or code terms. Code / source terms: Dataset license not stated on the reviewed paper page Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| TVL: Touch, Vision, and LanguageUse the authors’ citation instructions on the primary paper or project page.Download route linked; files not verifiedAccess checked 2026-08-19The official project page links the paper, code, dataset, and models.Reviewed 2026-08-19 | UC Berkeley; Meta AI Research; TU Dresden; CeTI 2024 | Robot: Handheld 3D-printed HCT collection device; Robot-collected SSVTP subset Sensor: DIGIT | Tactile RGB image; RGB image; English tactile description Reported / previously documented scale: 43,741 in-contact image-touch pairs; HCT contributes 39,154 pairs and SSVTP contributes 4,587 pairs Public-file audit: Unknown: no independently recounted public-file total in this record. Independent check: Existing source review only. No new dataset-file download, checksum validation, training run, or hardware test in this audit. | Tactile-semantic classification; Touch-vision alignment; Touch-language alignment; Tactile description generation Objects: In-the-wild surfaces and objects plus the SSVTP laboratory subset; a total object-category inventory is not stated on the reviewed project page. | Temporally aligned touch, vision, and open-vocabulary language examples; 10% of labels are human annotations and 90% are VLM-generated in the paper. Dataset-file license: Unknown: a dataset-file license has not been verified separately from paper or code terms. Code / source terms: Dataset license not stated on the reviewed project page Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| ObjectFolder RealUse the authors’ citation instructions on the primary paper or project page.Download route linked; files not verifiedAccess checked 2026-08-19The official ObjectFolder site provides dataset and benchmark download routes.Reviewed 2026-08-19 | Stanford University; Carnegie Mellon University 2023 | Robot: Franka Emika Panda Sensor: GelSight robotic finger | 3D mesh; HD RGB video; Impact audio; Tactile reading Reported / previously documented scale: 100 real-world household objects Public-file audit: Unknown: no independently recounted public-file total in this record. Independent check: Existing source review only. No new dataset-file download, checksum validation, training run, or hardware test in this audit. | Cross-sensory retrieval; Contact localization; Material classification; Reconstruction; Manipulation benchmarks Objects: One hundred real-world household objects; detailed inventory is provided through the official ObjectFolder resources. | Multisensory measurements aligned to object surface locations, including mesh, visual, acoustic, and tactile assets. Dataset-file license: Unknown: a dataset-file license has not been verified separately from paper or code terms. Code / source terms: Review the license bundled with the downloaded ObjectFolder release Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| ObjectFolder 2.0Use the authors’ citation instructions on the primary paper or project page.Download route linked; files not verifiedAccess checked 2026-08-19The official download page provides the Object Files, rendering instructions, metadata, and license information.Reviewed 2026-08-19 | Stanford University; Carnegie Mellon University 2022 | Robot: Simulation / neural object rendering Sensor: Simulated GelSight-style tactile rendering | Visual appearance; Impact audio; Tactile RGB image; 3D geometry Reported / previously documented scale: 1,000 multisensory neural objects Public-file audit: Unknown: no independently recounted public-file total in this record. Independent check: Existing source review only. No new dataset-file download, checksum validation, training run, or hardware test in this audit. | Object scale estimation; Contact localization; Shape reconstruction; Multisensory representation learning Objects: One thousand neural objects derived from real-object 3D assets across common household categories. | Object File implicit neural representations with query parameters for view, lighting, impact, contact location, gel rotation, and indentation depth. Dataset-file license: CC BY 4.0 for ObjectFolder 2.0 according to the existing review of the official download page. Code / source terms: CC BY 4.0 for ObjectFolder 2.0 according to the official download page Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| TacVerseUse the authors’ citation instructions on the primary paper or project page.Registration / approval requiredAccess checked 2026-09-13Official code and gated Hugging Face task archives verified on 2026-09-11. The 106,800-image count is paper-reported; files were not downloaded or recounted. No root code license was identified at the reviewed revision.Reviewed 2026-09-11 | Imperial College London; Queen Mary University of London; King’s College London 2026 | Robot: Controlled tactile data-collection platform Sensor: GelSightNoMarker; GelSightMarker; MagicGripper; MagicTac; TacTip; ViTac; ViTacTip | Tactile RGB image; Shape label; Grating label; Force label Reported / previously documented scale: 106,800 tactile images from seven vision-based tactile sensors Public-file audit: 7 paths in the provider file manifest, including metadata. Bytes, frames and episodes were not independently recounted. Independent check: Official hosting API and file manifest inspected at the pinned revision. No dataset payloads downloaded; no checksums, schema validation or training run. | Shape classification; Grating classification; Force regression; Zero-shot cross-sensor transfer; Few-shot adaptation Objects: Controlled shape and grating stimuli plus force-regression contacts; a general object-category inventory is not stated. | Multi-sensor tactile images and task labels organized for within-sensor, zero-shot cross-sensor, and few-shot protocols. Dataset-file license: cc-by-4.0 in the provider dataset-card metadata at the pinned revision; individual asset terms were not inspected. Code / source terms: CC BY 4.0 in the official Hugging Face metadata; access is gated and asset-specific terms were not independently checked Train / test split: Unknown: no train/test split manifest verified in this audit. | |
| VTDexManipUse the authors’ citation instructions on the primary paper or project page.Download route linked; files not verifiedAccess checked 2026-08-19The author project and repository link benchmark code, paper, and dataset access.Reviewed 2026-08-19 | Zhejiang University 2025 | Robot: Human dexterous demonstration platform; Simulated dexterous manipulation benchmark Sensor: Sparse binary tactile sensing | RGB image; Sparse binary touch; Proprioception; Manipulation trajectory Reported / previously documented scale: 10 daily manipulation tasks across 182 objects; trajectory and frame totals are not stated as text on the reviewed project page Public-file audit: Unknown: no independently recounted public-file total in this record. Independent check: Existing source review only. No new dataset-file download, checksum validation, training run, or hardware test in this audit. | Visual-tactile pretraining; Bottle-cap turning; Faucet screwing; Lever sliding; Table reorientation; In-hand reorientation; Bimanual hand-over Objects: One hundred eighty-two objects used across ten daily manipulation tasks. | Visual-tactile pretraining data and an Isaac Gym policy benchmark; inspect the official repository for the current file layout. Dataset-file license: Unknown: a dataset-file license has not been verified separately from paper or code terms. Code / source terms: MIT License for the benchmark code; verify dataset-file terms separately Train / test split: Unknown: no train/test split manifest verified in this audit. |
Short answer
What you need to know
- 1
A useful tactile dataset is defined by more than frame count. Check the physical collection event, sensor and robot state alignment, object and task diversity, split unit, access terms, and downstream evaluation.
- 2
Contact sequences matter because adjacent tactile frames from the same press or trajectory are strongly related. Random frame splits can leak near-duplicate contact evidence into both training and test sets.
- 3
Use the searchable directory below to find resources such as TactiDex, VTDexManip, EgoTouch, and Touch and Go. Filter by sensor, robot, task, or modality, then open the primary source to check downloads and reuse terms. A listed resource may still have incomplete files or an unstated license.
Topic 01
Does a paper describe data you can actually obtain?
A result table can justify reading a method without establishing a usable download. SlipSense describes a labeled slip collection; Touch2Trace describes pretraining and cable demonstrations; Visible Touch describes real-robot demonstrations and magnetic hardware. The original papers and available official project links checked on September 19, 2026 did not establish verified downloads and reuse licenses for these collections. They remain method references rather than new downloadable dataset records.
When evaluating a release, distinguish the observation data from its task labels, action contract and pretrained weights. For slip, request independent onset labels and object-level splits. For cable policies, request ordered observations, joint commands and excluded-start rules. For overlays, request raw readings, sensor geometry, camera transforms and normalization statistics. Unknown fields remain unknown until the released assets support them.
Topic 02
How to read this directory
The entries below are research resources with different goals; they are not interchangeable rows in one leaderboard. Some emphasize tactile-language and material understanding, others whole-hand contact, data collection, humanoid action alignment, or multisensory representation learning.
Before use, open the primary source and project page. Verify the actual downloadable files, license, sensor hardware, collection protocol, annotations, train-test splits, and version. A paper saying that a resource is open does not replace checking the current repository terms.
Topic 03
From a dataset shortlist to a loading plan
Start with the task and collection setup. Human demonstrations can support perception or representation learning without providing an executable robot action contract. Simulated touch can support controlled experiments without establishing calibration or transfer to a physical sensor.
Select up to three records in the explorer, then export the comparison. The CSV preserves the recorded availability, license, source-review date, primary paper, and official links. It is a snapshot of directory notes, not a redistribution of the underlying datasets or a grant of reuse rights.
Use the downloadable selection checklist to document sensor compatibility, action semantics, synchronization, splits, current file access, reuse terms, and a minimum loading run. Write down what you actually downloaded and decoded. Leave unresolved fields explicit instead of treating a project announcement as a ready-to-train release.
- Define the required observations, labels, actions, and target robot before choosing a resource.
- Inspect one complete sequence: units, timestamps, coordinate frames, missing values, and action meaning.
- Check the official host and dataset-file terms separately from the paper and code licenses.
- Record the version and review date; use the directory citation for these notes and cite the original authors for their research.
Topic 04
Tactile dataset and resource comparison
This comparison records the main research unit and the limit a user should preserve. Counts are included only where the primary source states them clearly.
| Resource | Signals and scale | Best-fit question | Evaluation unit | Evidence boundary |
|---|---|---|---|---|
| PRISM | 5,000+ robot trajectories, 5,000 paired human demonstrations, 45+ hours, and approximately 27M images across visual and visuotactile streams; tactile sensing covers an unspecified subset. | Industrial contact-rich manipulation across multiple robots, grippers, and teleoperation interfaces. | Split by complete episode, task, operator, hardware configuration, object, and modality availability. | The paper says open-sourced, but the official project still marks the dataset “soon”; no public data files or dataset-file license were verified on 2026-08-22. |
| EgoTouch | 208 human bimanual manipulation tasks across 1,891 episodes with head and dual-wrist RGB, 3D hand pose, and wearable pressure maps; the repository reports about 2M frames. | Egocentric vision-to-touch prediction and human contact representation pretraining. | Split by complete episode, task, object, scene, and subject or collection condition where exposed. | This is human-wearable interaction data, not robot action trajectories. Hosted files exist, but the official repository warns that upload may be incomplete and a separate dataset-file license was not verified. |
| HT-Bench | 10M egocentric RGB frames and 7.8M full-hand tactile frames collected across 226 tasks. | Full-hand tactile representation learning, cross-modal alignment, and unseen-task evaluation. | Split by task and trajectory; test held-out tasks, objects, sensor units, or embodiments for the claimed transfer. | A 2026 preprint. Scale does not make temporally adjacent frames independent, and reported results are specific to the benchmark setup. |
| RCT | 29,279 tactile frames from 122 industrial reference materials in 7 categories, collected with 3 DIGIT sensors; paired touch, image, language, and force context. | Material understanding and tactile-language retrieval. | Keep full press or contact sequences together; test held-out materials where possible. | A 2026 preprint. Reported performance is specific to its sensors, materials, models, and splits. |
| TactiDex | Whole-hand tactile observations aligned with multi-granularity kinematic and object states for single-hand and bimanual tasks. | Contact-rich dexterity and transfer across manipulation settings. | Use the standardized task and transfer protocol described by the source. | A 2026 preprint and project resource; inspect the released tasks, files, and license before reuse. |
| FreeTacMan | Paired visuo-tactile observations and interaction trajectories collected with a portable, human-operated workflow. | Scaling contact-rich demonstrations without occupying a robot arm for every collection session. | Split by task, object, trajectory, and operator conditions that match the transfer claim. | A 2025 preprint; human-device data still needs validation on the target robot embodiment. |
| Humanoid visual-tactile-action dataset | Synchronized vision, tactile observations, and action context for humanoid contact-rich manipulation. | Learning policies that need touch aligned with the action that produced it. | Keep synchronized trajectory segments and embodiment conditions intact. | A preprint; transfer depends on robot geometry, sensor placement, action space, and task distribution. |
| Sparsh-X research resource | Digit 360 tactile images, audio, motion, and pressure used for self-supervised multisensory touch representations. | Learning reusable tactile features across physical-property and manipulation tasks. | Evaluate downstream tasks and held-out conditions, not only pretraining loss. | A 2025 preprint tied to a multisensory sensor stack; cross-sensor transfer still requires evidence. |
| Touch and Go | Approximately 13.9K detected touches across indoor and outdoor scenes with egocentric vision and GelSight recordings. | In-the-wild visuo-tactile representation learning and future-touch prediction. | Split by object, scene, video, or collection sequence for the intended claim. | The object-instance count is estimated in the paper and the collection is human-operated rather than robot-action data. |
| TVL | 43,741 in-contact image-touch pairs with English tactile descriptions. | Touch-vision-language alignment and tactile description. | Keep HCT and SSVTP origins, contact events, and human versus generated labels visible. | The paper reports that 90% of labels are VLM-generated and documents occasional label errors. |
| ObjectFolder Real / 2.0 | 100 real household objects plus 1,000 neural objects with visual, acoustic, tactile, and geometric data. | Multisensory object recognition, reconstruction, and manipulation. | Separate real measurements from simulated neural-object rendering. | Synthetic and real results are not interchangeable; use the exact release and task protocol. |
| TacVerse | 106,800 paper-reported tactile images from seven vision-based tactile sensors. | Within-sensor learning, zero-shot transfer, and force-only few-shot adaptation. | The paper specifies whole-trial chronological 60/20/20 splits; sample_id-to-trial mapping remains unverified. | Official code and gated Hugging Face archives verified on 2026-09-11; raw data not downloaded or independently recounted. |
| VTDexManip | Visual-tactile data from 10 daily manipulation tasks across 182 objects plus a six-task dexterous benchmark. | Visual-tactile pretraining and policy evaluation. | Use task, object, modality, and policy splits that match the transfer claim. | The official code is MIT-licensed; dataset-file rights need separate verification. |
Topic 05
Why contact-sequence splits matter
RCT reports 29,279 frames but also preserves full contact sequences. Frames from one press are correlated, so a random frame-level split can place nearly the same physical event in training and test data. The paper reports that removing contact-sequence overlap reduces tactile-to-text Recall@1 by 17.7 percentage points.
The broader rule is to split at the level of the claim. For unseen-material performance, hold out materials. For unseen-object manipulation, hold out objects. For transfer across robots or sensors, hold out the target hardware. A large test set is not independent if the same contact event, object instance, or trajectory appears on both sides.
- Record the physical unit: press, grasp, trajectory, object, task, operator, robot, and sensor
- Group correlated frames before creating train, validation, and test sets
- Publish split manifests or deterministic split code
- Report results for the hardest held-out condition relevant to the claim
Topic 06
Dataset selection checklist
Start from the deployment mismatch you need to measure. A material dataset may be rich enough for tactile-language learning but unsuitable for action-conditioned robot control. A whole-hand trajectory dataset may support dexterity research but still mismatch a fingertip sensor, gripper geometry, or action space.
- Access: files, repository status, license, citation, version, and checksum
- Hardware: sensor model, serial variation, placement, sampling rate, calibration, and units
- Alignment: timestamps for touch, vision, force, pose, joint state, and actions
- Coverage: objects, materials, tasks, contact types, operators, robots, and disturbances
- Splits: sequence, object, material, task, sensor, or embodiment independence
- Outcome: retrieval, classification, prediction, imitation, or real-robot task success
Topic 07
What dataset size does not prove
Frame count does not establish diversity, independent evaluation, target-robot transfer, or production readiness. Ten thousand adjacent frames from a small set of presses can contain less independent evidence than a smaller collection spread across objects, sensors, and trajectories.
Use this directory to locate sources, then document the exact dataset version and split used in your experiment. Do not compare headline metrics across resources unless the sensor inputs, tasks, models, and evaluation protocols are genuinely aligned.
Topic 08
Inspect a dataset before adapting it for training
After checking access and license at the original source, inspect episode boundaries, feature meanings, timestamp conventions and missing observations. Container names do not guarantee compatible actions, camera timing or tactile inputs.
The LeRobot tutorial explains v3 storage and supplies a small numeric checker with valid and broken synthetic fixtures. It complements the tactile CSV exercise; neither is an original research dataset or a complete training-readiness certificate.
Recomputable directory analysis · v2026.09.19
Public availability and reproduction conditions
Compiled 2026-09-19 from the 20 records in this tactile directory. This is a selected research sample, not an industry census. Individual source-review and access-check dates appear in each record. A public manifest is evidence of hosted paths, not proof that all payloads download or a result can be reproduced.
| Evidence category | Records | Interpretation |
|---|---|---|
| Public file manifest | 7 / 20 | As recorded on each access-check date; access can change. |
| Announced; release not verified | 3 / 20 | As recorded on each access-check date; access can change. |
| Download route linked; files not verified | 6 / 20 | As recorded on each access-check date; access can change. |
| Access unknown | 3 / 20 | As recorded on each access-check date; access can change. |
| Registration / approval required | 1 / 20 | As recorded on each access-check date; access can change. |
| Dataset-specific license documented | 8 / 20 | 12 unknown; a paper or code license does not fill this field. |
| Split files listed | 1 / 20 | Listing only; split membership and leakage not checked. |
| Split tools / protocol described | 2 / 20 | Distinct from inspecting released splits; 17 records remain unknown. |
The practical constraint is the evidence gap between a paper-scale claim and a usable training package. This audit inspected 8 provider manifests without downloading dataset payloads. Start by pinning a release, reading dataset-file terms, checking the schema and units, and validating the train/test boundary before running a baseline.
Specific evidence still to verify
- Bench2Dex simulation demonstrations: dataset-file terms; payload completeness and reproduction remain unchecked.
- UniVTAC Encoder Pretraining Corpus: train/test split evidence; payload completeness and reproduction remain unchecked.
- UniVTAC Benchmark Dataset: train/test split evidence; payload completeness and reproduction remain unchecked.
- T-Rex Tactile-Reactive Dexterous Manipulation Dataset: train/test split evidence; payload completeness and reproduction remain unchecked.
- EgoTouch: dataset-file terms; payload completeness and reproduction remain unchecked.
- PRISM: Precision and contact-rich Real-world Industrial Skill dataset with Multimodal sensing: dataset-file terms; train/test split evidence; payload completeness and reproduction remain unchecked.
- SoftVTBench: train/test split evidence; payload completeness and reproduction remain unchecked.
- RoboTacDex: A Dexterous Visual-Tactile-Action Dataset for Humanoid Manipulation: dataset-file terms; train/test split evidence; payload completeness and reproduction remain unchecked.
- HT-Bench: dataset-file terms; train/test split evidence; payload completeness and reproduction remain unchecked.
- TactiDex: dataset-file terms; train/test split evidence; payload completeness and reproduction remain unchecked.
- FreeTacMan: train/test split evidence; payload completeness and reproduction remain unchecked.
- Humanoid Visual-Tactile-Action Dataset: dataset-file terms; train/test split evidence; payload completeness and reproduction remain unchecked.
- Sparsh-X Multisensory Touch Resource: dataset-file terms; train/test split evidence; payload completeness and reproduction remain unchecked.
- Touch and Go: dataset-file terms; train/test split evidence; payload completeness and reproduction remain unchecked.
- TVL: Touch, Vision, and Language: dataset-file terms; train/test split evidence; payload completeness and reproduction remain unchecked.
- ObjectFolder Real: dataset-file terms; train/test split evidence; payload completeness and reproduction remain unchecked.
- ObjectFolder 2.0: train/test split evidence; payload completeness and reproduction remain unchecked.
- TacVerse: train/test split evidence; payload completeness and reproduction remain unchecked.
- VTDexManip: dataset-file terms; train/test split evidence; payload completeness and reproduction remain unchecked.
Directory change log
— Added Bench2Dex simulation demonstrations with a pinned public manifest, separate license fields and original authors. STAR remains a paper-linked candidate outside the catalog count and Dataset schema. Published a new audit snapshot; older records retain their own verification dates.
— Added separate access evidence, provider revisions, public-manifest scale, dataset-file license status and split evidence to existing records, with computable counts and JSON export. Original audit snapshot.
Paper-associated data
Data candidates: access still unverified
These papers describe research data, but we have not verified a downloadable release and its reuse terms. Candidates are separate from the dataset catalog, its counts and filters, and its Dataset structured data. The JSON export preserves them in a separate candidates field.
STAR paper-associated dexterous manipulation data
Paper-linked candidate · Access checked 2026-09-18
Paper-linked data candidate only. Official project reopened on 2026-09-18; no dataset download, code release or dataset reuse license verified. Excluded from the dataset catalog count, available-file filters and Dataset structured data.
- Original authors
- Xiangcheng Liu; Tianhao Wu; Le Zheng; Yidong Wang; Bowen Jiang; Mingjie Pan; Xinlin Ren; Yi Liu; Jianlan Luo
- Source version
- First submitted 2026-09-11; arXiv v1 (2609.12549v1)
- Institutions / data origin
- Shanghai Innovation Institute; Agibot / real
- Reported scale
- Author-reported 200 hours, 10,576 trajectories and 65 tasks. No hosted-file count verified.
- Robot and sensors
- Bimanual wheeled mobile robot; two hands with 10 actuated DoF and 16 total DoF each; commercial model name unknown; 268-dimensional piezoresistive tactile sensing per hand; Two wrist RGB cameras and one head RGB camera
- Tasks and objects
- Dexterous multi-finger manipulation; Pick-and-place and insertion; Long-horizon manipulation. Task-specific household and manipulation objects; exhaustive inventory and unique object count unknown.
- Modalities, actions and synchronization
- Three-view RGB; Bilateral tactile signals; Robot state; Command actions; Task-level language instructions. Paper describes timestamp alignment to 30 Hz; this is not the raw tactile sampling rate. Command actions use end-effector poses and hand joints (34 active dimensions, padded to 64 for the policy). Public package, file schema and train/validation/test manifests unknown.
- Splits and generalization
- Evaluation objects are held out from 200-hour pretraining but used in 100 task-specific post-training trajectories per task; 20 evaluation trials per task use held-out initial configurations. Not unseen-object zero-shot generalization.
- Code / data / model / asset terms
- Unknown: no dataset-file, code, weight or asset license verified. The arXiv article license does not license the described dataset.
- Verified and unresolved
- Paper v1 and official project inspected. Download entry, released file inventory, payload integrity, split files and dataset license remain unknown. No independent reproduction by RoboSkin.
Paper routes
Start with source-backed RoboSkin briefs
Tactile datasets / 2026-09-11TacVerse tests whether touch transfers between sensorsTacVerse compares shape, grating, and force perception across tactile sensors. Its public code and gated dataset are now identifiable, but transfer quality depends strongly on the task.Tactile datasets / 2026-08-22PRISM maps 5,000+ contact-rich industrial robot trajectoriesPRISM reports 5,000+ robot trajectories across 25+ industrial tasks, but tactile observations cover only a subset and the announced dataset download is not yet available.
Tactile AI / 2026-08-22HT-Bench full-hand tactile benchmark for robot manipulationHT-Bench v2 pairs egocentric vision with millions of full-hand tactile frames, corrects the vision-to-tactile metric split, and adds four real-robot evaluations.
Tactile Data / 2026-08-22FreeTacMan robot-free visuo-tactile data collection for tactile AIA research note on FreeTacMan, robot-free visuo-tactile datasets, tactile AI data collection, and why robot skin models need contact diversity.
Tactile Data / 2026-08-22GIST humanoid visual-tactile-action dataset maps 101.9K soft-object samplesThe GIST preprint reports 101.9K visual-tactile-action samples for towel and sponge manipulation, with dense hand touch, two camera views, and explicit access limits.Common questions
FAQ for this topic
What is the best tactile dataset for robot learning?
There is no universal best dataset. Match the resource to the learning goal, sensor signals, robot embodiment, task, and evaluation split required by the deployment claim.
Why are random frame splits risky for tactile data?
Adjacent frames from the same press, grasp, or trajectory can be near duplicates. If they appear in both training and test sets, the metric can overstate generalization.
Does open source mean unrestricted commercial use?
No. Open access to a paper, project page, code, or files does not define commercial rights. Check the license for the exact dataset version and every bundled asset.
Should tactile datasets include robot actions?
They should when the goal is policy learning, action-conditioned prediction, or replay of manipulation. Material recognition or representation learning may use different labels and collection units.