Short answer
Tactile data records contact from skin-like or fingertip sensors: pressure maps, shear, vibration or deformation images. It matters for in-hand manipulation, slip, and handling fragile or deformable objects where vision cannot see the contact patch. Tactile data is tied to the sensor that recorded it, so buyers should check what exists for their hardware and expect to treat it as a specialist collection with strict sensor, calibration and synchronization specifications.
The buyer's problem
Humanoid hands will hold objects they occlude. A camera cannot see slip beginning under a fingertip, or how firmly a cup is held. Tactile streams carry that information, but before relying on any dataset a buyer should check how large it is, which sensor recorded it and whether it transfers to their hardware. A buyer asking "where can I buy tactile data" is usually choosing between existing research datasets, specialist labs that run tactile rigs, and a commissioned collection on their own hand.
What published datasets include
RH20T is one reference for contact-rich data: more than 110,000 real-world manipulation sequences, each with visual, force, audio and action data, on multiple robots and camera viewpoints, with each robot sequence paired to a human demonstration video and a language description (Fang et al., 2023). It records force rather than describing tactile skin data, so use it as an example of multimodal contact-rich collection, and check the paper for the exact sensors before assuming anything about touch.
The wider lesson from published collection systems is about what is absent. DexCap, a portable hand motion-capture rig that combines SLAM with electromagnetic-field tracking, records 3D scene observations and was evaluated on six dexterous tasks, but it provides no force sensing and has a roughly 40-minute session limit set by power (Wang et al., 2024). Hand pose from a mocap glove or rig tells you where fingers were, not how hard they pressed. Tactile data is a separate channel that someone has to instrument and record.
Tactile modalities to distinguish
| Modality | What it records | Check before buying |
|---|---|---|
| Pressure array or taxel grid | Normal force per cell | Grid size, units, sampling rate, dead cells |
| Vision-based tactile | Camera images of a deforming gel | Gel wear, lighting, image calibration, marker pattern |
| Shear and slip signals | Tangential force or vibration | Sensor bandwidth, filter settings |
| Whole-hand or skin coverage | Contact over palm and links | Layout map, calibration per sensor |
| Joint torque estimates | Inferred contact from motors | Not a tactile sensor; state clearly |
A supplier who says "tactile" should name the sensor type, vendor, mounting location, and the raw unit. If the data is processed (contact masks, estimated forces), ask for the raw stream too.
Why transfer between hands is hard
Tactile signals depend on the sensor's material, geometry, wear and mounting. A dataset recorded with one sensor on one finger shape does not directly drive a different hand. Possible bridges include converting to a common representation such as contact location and estimated force, or training an encoder per sensor with a shared task head. These are research approaches; none is a purchase specification. Buyers who need data for a specific humanoid hand should expect to collect on that hand or an identical sensor stack. Tactile calibration sits beside camera and hand-eye calibration in calibration requirements. Retargeting visual hand motion to a robot hand is a different issue, covered in retargeting human hand demonstrations.
What tactile data cannot support
Tactile logs do not provide scene context, task language or objects' global pose unless recorded with cameras. They say little about what the robot should do next without paired actions. Sensor drift, hysteresis and temperature effects can change readings across a session. Check whether a public tactile set was recorded at tabletop or lab scale; workplace realism would require on-site collection with robot hardware. Human workplace video has no touch channel, which is why video alone is not enough for contact-rich skills. Business archives and SOPs do not contain tactile data.
Illustrative specification for a tactile collection
# Illustrative - example values only; buyer sets every value
tactile_spec:
sensors:
- {id: tip_L1, type: vision_based_gel, mount: left_index_fingertip, frame_rate_hz: 60}
- {id: palm_L, type: taxel_array, grid: [8, 8], sample_rate_hz: 200}
raw_streams_delivered: true
derived_streams: [contact_mask, estimated_normal_force]
calibration:
per_sensor_report: required
no_contact_baseline_per_session: required
reference_load_test: "known weights at start and end"
sync:
reference_clock: shared_with_joint_states
max_offset_ms: 5
episode_metadata: [object_id, grasp_type, slip_event_flags, sensor_wear_count]
storage: mcap_or_hdf5
Decisions to make before commissioning
Before commissioning, settle what contact information the policy will consume. A policy that reads a contact mask needs different data from one that reads dense pressure frames or a force estimate. Decide whether tactile input is used at training time only, for example to label contact events, or also at deployment, because a sensor that is absent on the deployed hand cannot be an input there. Decide how many sensor replacements the collection can tolerate, since swapping a gel or skin mid-campaign changes the signal. Record the object set, because stiffness, texture and weight shape what the sensor sees, and keep a log of which object and which grasp each episode contains.
Acceptance checks
- Baseline stability. With no contact, readings stay within a limit you set across the session.
- Reference load. A known mass produces a repeatable response, before and after the session.
- Dead-element report. Count of failed taxels or damaged gels per sensor.
- Sync test. Contact onset in the tactile stream aligns with gripper closure and visible touch within your tolerance. See sensor synchronization requirements.
- Label audit. Sample slip or contact event flags and compare them with video.
- Sensor change log. Replaced or recalibrated sensors are recorded by date.
The general delivery process is in acceptance testing for deliveries.
Sourcing routes
Research labs and specialist sensing providers are the places to check for tactile supply; ask any provider whether a commercial catalog exists or whether data is collected only to order. Existing sets may carry academic licenses, so check public dataset commercial license checks before planning a product around one. Commissioned collection allows you to specify the hand, objects and tasks, at the cost of time, hardware access and iteration. Force at the wrist may be a cheaper, more portable substitute for some tasks; see force/torque data for contact-rich tasks. Route comparison is in existing assets, new collection or specialist providers.
For rights, ask whether the data records only robots and objects or also people; if no human wore the sensors, personal-data questions may be limited, but licensing terms still apply. If human demonstrators wear gloves or sensors, performer releases apply.
Questions to ask suppliers
- What exact sensor model, mounting location and sampling rate were used?
- Is the raw sensor output delivered, or only processed contact estimates?
- How was each sensor calibrated, and how often?
- How do readings change with sensor wear, temperature or repeated contact?
- Which hand or gripper carried the sensors, and can we replicate the stack?
- How are tactile streams synchronized with joint states and cameras?
- What objects, materials and grasp types are covered, and what is missing?
- Are slip or contact events labeled, and who checked the labels?
- Is any human wearing sensors or gloves, and are releases in place?
- What license terms apply to commercial training and to derived models?
Discuss your requirements
If you need touch data for a specific hand, describe the sensor stack, tasks and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates license and delivery; if no existing asset fits, SourceX can scope a commissioned collection. A request does not guarantee a match. Start at the humanoid training data guide or the procurement hub.
Discuss your humanoid training-data requirements with SourceX
Sources
- RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shot — RH20T pairs visual, force, audio and action data per contact-rich sequence.
- DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation — DexCap records hand motion capture without force sensing.