Rabbit AI Request access

Prototype · early access

Every hour
usable and
genuine.

Capture infrastructure for robot training data. Checks on the phone tell collectors to re-record before junk is uploaded. Every segment is signed and fingerprinted, so duplicated, recycled or re-filmed footage is caught before anyone pays for it.

Become a design partner How it works
REC session 7f3a·c219 00:00:24
cam0 + imu · one clock · hands ✓ framing ✓ sync ✓
signed hash chain · 2 s segments
seg 0011 · 9b2e…44 ← 3f9c…a1 · signed 12 verified0 issues
  • Checked on the phonehands, framing and sync, while recording
  • Signed at capturea tamper-evident chain of segments
  • Duplicates caughtacross vendors, even after re-encoding
  • Usable-hours reportwith every delivery

The problem

Labs pay for hours they can't use.

Robot-learning labs buy hours of human demonstration footage. Much of what gets recorded is never usable, or isn't genuine, and the waste is found late: after collectors are paid, after delivery.

  • Unusable

    Filtered out

    Hands out of view, bad framing, desynced sensors, broken uploads. Labs drop these hours, and more of them now pay only for usable data.

  • Not genuine

    Paid for twice

    Duplicates, recycled public footage, undisclosed resale, and clips re-filmed or gamed by open contributor networks.

  • Today

    Rebuilt everywhere

    Every vendor builds its own capture app, upload, QA and anti-fraud. Labs re-check everything after they've paid.

How it works

Checked and signed before it leaves the phone.

  1. 01

    Capture + check

    Video and motion recorded in 2-second segments on one device clock. Checks on the phone flag hands out of view, bad framing or broken sync, so the collector re-records on the spot.

    hands ✓ · framing ✓ · sync ✓
  2. 02

    Sign

    Each segment is hashed and chained to the one before, designed to sign with the phone's secure hardware. Consent and device details are bound into the record.

    seg 0007 · prev → 9b2e…44 · sig
  3. 03

    Match

    Each segment's video and motion are fingerprinted and checked across vendors and against public datasets, so a re-encoded or recycled copy still matches. Private by default.

    video phash · imu fingerprint
  4. 04

    Deliver + report

    Synced streams upload resumably in lab-ready formats, with a usable-hours report: what passed, what was flagged, and why. Labs can verify it themselves.

    usable · unique · flagged
Steps 1–2 · on the phone, in our capture app or SDK Steps 3–4 · run by us, neutral across vendors

Footage is encrypted and uploaded directly, so it never lands in the collector's camera roll.

What it catches

Usable and genuine hours.

An hour counts only if a lab can train on it and it is what it claims to be. We check both, at capture, and the verifier shows exactly what failed.

Usable

Footage a lab can train on.

  • Hands out of view or out of frame.
  • Bad framing or a blocked camera.
  • Video and motion out of sync, or a sensor stream dropped.
  • Broken or partial uploads.

On-device quality checks · in development

Genuine

Footage that is what it claims.

  • Cut, reordered or edited segments, and motion swapped in from another session.
  • Duplicates and re-encoded copies, across vendors and against public datasets.
  • Footage resold or recycled without disclosure.
  • A screen re-filmed instead of real work.

Tamper + duplicate checks working on test data · re-film check in development

attack · cut-5sFAIL
Consent
signed + bound
Content vs hashes
match
Hash chain
broken · 4 segments with problems
FAIL · 11/15 segments OK
cam0: 6.0 s missing: seq 5–7, t = 00:10–00:16
cam0: chain break at seq 8

A 5-second cut, located.

Segments are 2 seconds long, so a 5-second cut removes three. The verifier names them and the exact break. No scrubbing through footage.

check a file · copy_reencoded.mp4DUPLICATE
C2PA manifest
none · stripped
Exact hash
no match · every byte changed
Fingerprint
match · signed origin 49277de7 · seg 0–14

A re-encoded copy, matched.

The copy lost its signed metadata and every byte changed. Its fingerprint still lines up with the signed original, frame by frame.

30/30tamper tests passed
0.21 sto verify a 30-minute session

On test data, the prototype passes 30 of 30 tamper tests and verifies a 30-minute session in 0.21 seconds.

Known limits The prototype runs on test data and signs with a software key. It doesn't yet run the on-device quality checks or catch a screen being re-filmed. We're building both: real-time checks for hands, framing and sync, and a check that compares the phone's motion data with the video.

Prototype

Working end to end.

Phone capture, signed hash chain, verifier and duplicate registry, running on test data today. On-device quality checks and the usable-hours report are next.

The prototype dashboard in dark mode for a finished recording that passed verification. A signed hash chain shows a genesis record followed by 20 verified two-second segments with 0 issues, 0 missing and 0 flagged. Below it, the recorded footage and a motion-sensor chart on the same clock.
Prototype dashboard · live session · test dataThe prototype signs with a software key today. It is designed to sign with the phone's secure hardware.

Who it's for

One usable-hours record, trusted on both sides.

Robot-data aggregators

Stop paying for hours you can't sell.

  • Run our capture app under your brand, or add the SDK to the app you already use.
  • Collectors are told to re-record on the spot, before junk is uploaded or paid for.
  • Catch duplicates, recycled footage and gamed clips before payout.
  • Ship every delivery with a usable-hours report the lab can check.

$0.50 per usable hour delivered, with volume tiers · free design-partner pilot, then a $1,000/month minimum

Robot-learning labs

Pay per usable hour, across every vendor.

  • Verify a delivery in seconds, before training.
  • One usable-hours report format across all your vendors.
  • Flag duplicates across vendors and public datasets, including footage resold as new.
  • Consent and device records attached, for legal review when you need them.

Verifier + usable-hours reports · subscription · early access for a small number of labs

Principles

Neutral. Plain about trust. We never sell data.

  • Neutral

    One stack and one registry across every vendor that joins. Each vendor makes it more useful to every lab.

  • No data business

    We never buy or sell data, so we don't compete with the aggregators or labs who rely on us.

  • Trust, labelled

    Our highest tier is our own signed capture app on hardware-attested phones. Every report says which tier each hour came from.

  • Consent and privacy

    Wearer consent is signed into the record and location is off by default. Duplicate matching is private and license-aware.

Team

Two technical founders.

Rithwick Sethi

Rithwick Sethi

Co-founder · CEO

  • MS, AI Engineering (ECE), Carnegie Mellon
  • Built firmware trace tooling at Apple
  • Autonomous drones: UAV swarm research
  • 2× Smart India Hackathon national winner
  • 7 published research papers

Leads the capture SDK and on-device signing.

LinkedIn ↗
Om Kulkarni

Om Kulkarni

Co-founder · CTO

  • MS, ECE and Engineering & Technology Innovation Management, Carnegie Mellon
  • Built production data pipelines at Wells Fargo
  • Benchmarked AI workloads across CPUs at Arm
  • Built robot capture and VLA architectures at CMU
  • Multimodal serving research at CMU Catalyst

Leads the backend, registry and verifier.

LinkedIn ↗
Based in

Pittsburgh, PA

Carnegie Mellon University.

Met at CMU in 2025 and co-founded NAAMI together.

Early access

Become a design partner.

We're looking for a small group of robot-data aggregators and robot-learning labs to pilot on one real collection project, free.

Become a design partner Book a call with the founders Request early access