Data infrastructure for physical AI

The tactile layer of physical AI.

Tactra captures how skilled human hands do real work — first-person, at industrial scale — and turns it into the training data the next generation of robots will learn from.

01 — Thesis

Robots have read the internet. They have never watched a welder.

Language models scaled on trillions of tokens of human text. Robot foundation models have no equivalent corpus: nobody has recorded, at scale, how people actually use their hands to make and fix things. That missing dataset is the bottleneck of physical AI.

Web text available to LLMs
1013tokens

Order of magnitude of text frontier language models are trained on.

Real manual-work video, public
≈0at scale

Existing egocentric datasets are small, academic, and mostly domestic — not skilled industrial work.

Hands doing skilled work, today
0

People worldwide whose daily labor is the richest untapped training signal for robotics.

02 — Platform

Vertically integrated, from lens to label.

Data quality is decided at capture. We plan to own every layer — so every hour that leaves Tactra is synchronized, consented, annotated and ready to train on.

L1

Capture hardware

rig

Head-mounted, wide-FOV RGB with IMU and optional depth — designed for eight-hour shifts, not lab sessions. Cheap enough to ship by the thousand.

L2

Operator network

people

Paid, consenting skilled workers in factories and workshops, recruited through local partners. Fair pay per hour of usable footage.

L3

Processing pipeline

compute

Ingest, sync and de-identify; then 3D hand pose, object tracks and task segmentation — automated, with humans in the loop for quality.

L4

Licensed corpus

product

Off-the-shelf dataset releases plus custom collection for labs with a specific bet — a task, a trade, a region.

03 — The record

One hour of work, fully legible.

Every clip is a structured record a robotics team can load and train on the same day: synchronized streams, hand and object geometry, and step-level task labels.

VideoEgocentric RGB · 1080p · 60 fps
Motion6-axis IMU · head pose
Hands21-keypoint 3D pose · both hands
SemanticsTask → step → action segmentation
ProvenanceConsent, location, trade, skill level
record.schema.json
// illustrative schema — planned format
{
  "clip_id": "tx_00f3a9c2",
  "trade": "wire_harness_assembly",
  "duration_s": 3600,
  "streams": {
    "rgb":  { "res": [1920, 1080], "fps": 60 },
    "imu":  { "hz": 400 },
    "hands":{ "keypoints": 21, "space": "3d_cam" }
  },
  "segments": [
    { "t": [12.4, 19.8], "step": "route_wire",  "tool": "none" },
    { "t": [19.8, 23.1], "step": "crimp_terminal", "tool": "crimper" }
  ],
  "operator": { "skill_yrs": 11, "consent": "signed", "pii": "removed" }
}
04 — Roadmap

Each release ten times the last.

A deliberate scale ladder: prove quality on a narrow slice of trades, then compound the collection network.

Tactra-1K1,000 hoursPhase I · pilot
Tactra-10K10,000 hoursPhase II
Tactra-100K100,000 hoursPhase III
Tactra-1M1,000,000 hoursPhase IV
Tactra-10M10,000,000 hoursHorizon

All figures are targets. Bars on log scale.

05 — Coverage

The work robots will do next.

We start where dexterity is hardest and demand for automation is highest — then widen the map.

T-01

Electronics assembly

PCB handling, soldering, inspection

T-02

Wire-harness

Routing, crimping, taping

T-03

Garment & textile

Cutting, sewing, folding

T-04

Metal fabrication

Welding, grinding, fitting

T-05

Maintenance & repair

Diagnostics, disassembly, refit

T-06

Food processing

Prep, portioning, packing

T-07

Logistics

Picking, kitting, palletizing

T-08

Craft & finishing

Woodwork, ceramics, leather

06 — Why now

The models are ready. The data isn't.

01

Foundation models went physical

Vision-language-action models now transfer across robots and tasks. Their ceiling is set by the breadth of real demonstrations they see.

02

Capture got cheap

Wearable cameras, IMUs and on-device compute cost a fraction of what they did five years ago — collection can finally scale with people, not labs.

03

Teleop doesn't scale

Robot teleoperation is slow and expensive. Human egocentric video is the only source that can reach millions of hours of skilled work.

07 — Team

Vision, AI, and the craft of scale.

CR

Carlo Rubinstein

Co-founder · AI strategy

AI strategist. Leads AI adoption and strategic projects inside a technology CTO office — turning frontier models into products and workflows that ship.

MK

Michael Kolomenkin, PhD

Co-founder · Computer vision

Technion PhD in 3D shape analysis and computer vision, published at CVPR. CTO of Fermata, building AI vision systems that run in the real world.

YK

Yaniv Kantor

Co-founder · Brand & growth

Founder and CEO of HYBRID, a Tel Aviv digital agency. Serial founder who has built and scaled brands, products and teams for over fifteen years.

We're raising

Help build the corpus robots learn from.

We're opening our first round to investors who believe physical AI will be won on data. Ask for the deck and we'll set up time with the founders.