Tactile data for physical AI

Your robot is flying blind on contact.

Every public dataset is vision-only demonstrations of things going right. Pisikal collects what the hand actually feels — and the failures, slips, and edge cases nobody else records.

GelSight · live
shear 0.42 N
normal 1.18 N
contact imprint elastomer · 24 fps
01 The modality gap

Vision sees the object.
Touch knows the grip.

A policy trained only on cameras has no signal for the moment that decides every manipulation task: contact. Whether the grasp holds, whether the object is slipping, how much force a soft thing can take. That information lives in the fingertips — and it's missing from almost every robotics dataset on earth.

Vision-only · the status quo

What cameras capture

Geometry, color, pose, motion. Enough to plan a reach. Blind the instant the gripper closes.

  • No contact force or slip signal
  • Occluded the moment of grasp
  • Can't tell rigid from deformable
  • Demonstrations of success only
Tactile + the tail · Pisikal

What touch adds

The contact event itself, captured as data: pressure fields, shear, slip onset, and the failures that teach robustness.

  • Normal & shear force at the fingertip
  • Slip detection before it drops
  • Deformable & fragile object handling
  • Edge cases and recoveries, on purpose
02 The long tail

Robots don't fail in the average.
They fail in the tail.

Nominal demos teach the easy 80%. The slip, the clutter, the occluded grab, the object that deforms — that's the 20% that breaks deployed policies.

And it's exactly the data nobody collects, because it's hard to stage and harder to label. We collect it on purpose: structured failure, edge cases, and recoveries, mapped to a taxonomy your team can train against.

nominal demos edge cases where robots fail
task-success distribution · the tail is unlabeled and uncollected
03 What we collect

Three things missing from your training set.

Tactile is the foundation. Edge cases are the proof. Custom is how it arrives shaped to your robot, your gripper, your task.

Core
/ 01 — TACTILE

Tactile signal

Pressure fields, normal and shear force, slip onset, and surface texture — captured at the fingertip with GelSight-class and force-torque sensing. The contact event as data.

/ 02 — EDGE CASE

Failure & edge cases

Slips, occlusions, clutter, deformables, bad lighting, and the recoveries that follow. Staged on purpose and labeled to a taxonomy — the tail your policy never saw.

/ 03 — CUSTOM

Custom to your task

Your embodiment, your gripper, your objects, your environment. We collect to your spec and annotation schema so the data drops straight into your training pipeline.

04 For the skeptics

What's actually in the data.

We're early — and direct about it. Here's the collection spec we're building design partners around. Modalities, taxonomy, and schema are co-defined with our first partners, not handed down.

Modalities

  • Tactile imprint GelSight-class
  • Force / torque 6-axis F/T
  • RGB-D multi-view
  • Proprioception joint + EE pose
  • Action labels teleop / kinesthetic

Edge-case taxonomy

  • Slip & regrasp contact
  • Occlusion visual
  • Clutter & collision scene
  • Deformable / fragile object
  • Lighting & reflectance sensor

Annotation

  • Contact event windows temporal
  • Success / failure / recovery outcome
  • Force regime tags per-frame
  • Object & material class semantic
  • Custom schema to partner

Delivery

  • Format LeRobot / HDF5
  • Embodiment your gripper
  • Licensing per-dataset
  • Collection operator network
  • Cadence design-partner

Early access: We're onboarding a small group of design partners to co-define modalities and taxonomy against real training needs — and to be first in line for the data.

05 Why now
There's no ImageNet for physical AI — and when it arrives, it won't be more demos. It'll be touch, and the failures.

Vision and language got their data flywheel a decade ago. Manipulation never did. The field is scaling demonstrations, but demonstrations of success in clean conditions don't transfer to messy rooms and fragile objects.

The bottleneck isn't model capacity anymore — it's the absence of contact-rich, failure-aware data. That's the gap Pisikal exists to fill, collected to your task instead of scraped from whatever happened to be lying around.

// Pisikal — physical AI training data

Get the data your robot
actually needs.

Tell us your embodiment and the tasks that keep breaking. We'll scope a tactile and edge-case collection around them.

Book a call →

Now onboarding design partners · labs & robotics teams