What Is Tactile Data in Robot Learning?

A camera shows a hand near an object, not how hard it presses or whether the object is slipping. Touch sensors read that directly, which is why contact-heavy, dexterous work is where buyers ask about tactile signals.

Firsthand’s standard capture lists no tactile or force stream. It uses cameras, depth and IMU streams, and contact_events is a label for when and where contact was made or broken. A separate glove tier records a tactile glove signal beside the video and is described on its own page. Touch sensor type, sample rate and units are Not stated for standard capture.

Tactile data is the signal from touch sensors. It describes the contact itself: where a hand or gripper is touching an object, how hard it presses, and whether the object is sliding or vibrating. The sensors usually sit in a robot fingertip, palm or skin, or in a glove worn by a person.

Vision records what the camera sees. Proprioception is a robot’s sense of its own body, such as joint positions and gripper state, and a person wearing a camera has none. A contact event is a label that says contact started or stopped. Tactile data is a separate measurement of the contact, and the only one of these that says how hard.

When a hand closes around an object, the fingers can hide the contact from a head camera. In its comparison of hand action representations, EgoScale (arXiv 2602.16710) reports that small errors in fingertip pose “often lead to implausible joint configurations after mapping, causing unstable grasps or contact loss in contact-sensitive tasks such as Cards and Bottle.”

That passage is about how hand motion is represented, not tactile sensing, and this page makes no claim that touch data would have fixed those failures.

The two papers read for this page describe vision and pose. EgoScale says each human demonstration “consists of egocentric RGB observations captured from a head-mounted camera, together with estimated camera motion and human hand pose.” EgoVerse (arXiv 2604.07607) says its data is processed into a unified format “that contains at minimum egocentric videos, hand keypoints, and camera poses.” Neither passage lists a tactile stream.

A glove in these papers can still mean pose rather than touch: EgoScale’s aligned human data uses gloves that “record full in-hand pose as 25 joint transforms.” These statements cover those passages only.

The contact-event page says Firsthand does not describe contact events as force, pressure or tactile data, and that standard capture uses cameras, depth and IMU streams, so contact events are not read from contact sensors. The hand-object interaction page lists force or tactile data as Not stated. A separate pilot-ready glove tier records a tactile glove signal beside the video; its own page covers it, and this page does not restate it.

What those pages do state is vision and pose: a head-mounted global-shutter camera, two wrist cameras, depth from a calibrated stereo rig, 21-joint 3D hand pose per hand with a visibility flag, and three 6-DoF IMUs at 200 Hz on one PTP clock. The second table lists each item and its status.

Custom collection starts from a written skill spec, and where coverage is already deep a signed spec reaches a first validated batch in a median of 48 hours. A request that needs touch should say so first, with the sensor, sample rate, units and sync it requires.

What is tactile data in robot learning?

It is the signal from touch sensors that measure contact between a hand or gripper and an object: where it touches, how hard it presses and whether it slips. It is a measurement, not a label or an image.

Is a contact event tactile data?

No. A contact event is a label for when and where contact was made or broken. Firsthand does not describe it as force, pressure or tactile data, and it does not say how hard anything was held.

Does egocentric video contain tactile data?

No. A head-mounted camera records what the wearer sees. Pressure and slip need touch sensors, so video alone carries no tactile signal.

Does Firsthand deliver force or tactile data?

Standard capture does not: it lists cameras, depth, 3D hand pose, contact event labels and IMU streams, and force or tactile data is Not stated there. A separate glove tier records a tactile glove signal; see its own page.

Can contact events or hand pose replace tactile data?

They answer different questions. Neither says how hard anything was held, and the site states no result for a policy trained with them in place of a touch signal.