2026 dataset directory
Tactile datasets for robot learning
Compare tactile datasets for robot learning by signals, collection unit, split design, task fit, access evidence, and transfer limits.
Published 2026-07-20 | Updated 2026-07-20 by RoboSkin.ai Editorial Team
- 5
- sections
- 4
- questions
- 6
- next routes
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
Choose the dataset that matches the intended learning problem: material understanding, whole-hand contact, imitation learning, multisensory representation learning, or target-robot control.
Topic 01
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 02
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 |
|---|---|---|---|---|
| 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. |
Topic 03
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 04
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 05
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.
Paper routes
Start with source-backed RoboSkin briefs
Tactile AI / 2026-06-18Sparsh-X multisensory touch representations for tactile AISparsh-X fuses image, audio, motion, and pressure from Digit 360, showing how multisensory touch can improve tactile AI for robot manipulation.
Tactile Data / 2026-06-18FreeTacMan 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 / 2026-06-18Humanoid visual-tactile-action dataset for contact-rich manipulationA research note on humanoid visual-tactile-action datasets, contact-rich manipulation, multimodal robot data, and why humanoid tactile learning needs synchronized action context.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.