Praxis

Y Combinator S26 · Robotics data infrastructure

A new 100% margin revenue stream from operations you already run.

Praxis turns the task footage your workforce generates every day into high-fidelity humanoid robotics training data — and pays you for it. Zero upfront cost, zero disruption to your crews.

Rendering earth

Collection network24 COUNTRIES

24

Countries engaged

01Who we are

A robotics company solving the bottleneck in front of every robotics company.

Praxis is a Y Combinator S26 robotics company building the world's largest network of real-task training data. We install lightweight capture hardware into workforces that already exist, annotate what they produce, and sell the resulting datasets to frontier AI labs — sharing the revenue with the businesses that host collection.

Y Combinator S26RoboticsIndustry agnosticCountry agnosticRevenue share

Operating footprint

We service frontier labs across warehouse, assembly, logistics, packaging and manufacturing sites, with live or contracted collection in 5 countries across 3 regions and 3,650 workers enrolled across 14 sites. A further 18 countries are in the pipeline, and utility and industrial field services are wide open.

More about the company and the team

The opportunity

Data is the biggest bottleneck in robotics.

01

Robots are trained on data

Humanoids learn from real task footage — egocentric first-person video, and teleoperation data captured alongside haptic gloves. Nothing about the current generation of models works without it.

02

Datasets don't exist at scale

Diverse, high-fidelity datasets for humanoid robotics are not available at the volume frontier labs need. The corpus that exists today is narrow, repetitive, and mostly synthetic.

03

Labs are paying for it now

Frontier AI labs are well capitalized and competing for the same scarce supply, bidding up the price of real task data. Demand is outpacing supply today, not eventually.

Task data demand is outpacing supply

What makes a dataset valuable

Three drivers set the price of every hour.

The more varied the data, the more valuable it becomes.

01

Where the data is collected

Country of origin is the single biggest driver of value. US data commands the most demand because of commercial pull and US-specific products and services.

02

Length & complexity of the task

Long, multi-step tasks teach a robot sequencing and judgment — not just a motion. Complexity compounds value.

03

Diversity of the dataset

Footage spanning geographies and industries generalizes. We are deliberately industry- and country-agnostic for exactly this reason.

02What we do

Three steps. Zero disruption to current operations.

Praxis converts latent task footage into annotated robotics training data and sells it into a market that is already bidding. Open any step for the detail.

Footage is priced, tagged, and paid within days

Capture to training data

What the footage looks like, and what it becomes.

The same seconds of one session, side by side — capture on the left, the structured output on the right.

Raw capture
Egocentric video with sensor gloves, recorded on a normal task. No staging, no retakes.
Annotated output
The same seconds after processing: per-hand keypoint tracking, motion signal, and labelled task segments.
Task organize bedroomSegment fold yellow pillowcase, place bottom right of bedSource 2:00.0 – 2:41.6 · 30 fpsShown at 2× speed

Input

Unstructured first-person video of real work

Process

30 fps annotation, PII redaction, quality control

Output

Structured, labelled task data licensed to frontier labs

03How we do it

Three hardware tiers, one pipeline, and a hard line around your IP.

Every tier is paid for and replaced by Praxis. Fidelity climbs with worker burden, so most partners start at tier one and move up on the task categories that justify it.

Paid for, shipped, and replaced by Praxis. Hardware plugs into any smartphone.

Model placeholderDrag to rotate

Tier 01

Smart glasses headmount

Egocentric video from a sub-10mm frame. Clears hard hats and arc-flash PPE, so it drops into regulated environments without a safety review of the work itself.

Capture
First-person video
Frame
Sub-10mm
PPE
Hard hat + arc-flash safe
Worker burden
Minimal
  • Best starting point for a first pilot — highest worker acceptance, fastest to deploy across a large crew.
  • Placeholder: drop the real headmount scan into /public/models and set `model` on this tier.

How data is collected, uploaded, annotated & secured

01

Collection

Workers record dexterous manipulation during normal jobs. Hardware plugs into any smartphone — iOS or Android.

02

Upload

Two routes, your choice: direct to cloud, or on-prem hard drives that never touch an outside network.

03

Annotation & QC

30 frames per second annotated by our research team using proprietary software, then quality-controlled before pricing.

We are not interested in trade secrets

We do not film trade secrets or IP, and we delete any footage on request — no questions, no retention window.

Privacy and security are enforced automatically

Faces, screens, license plates, and documents are detected and blurred or removed before storage. PII redaction is automatic, not a manual pass.

Fully customizable to your operation

Collection scope, task types, and reporting cadence are tailored to match how your sites already run.

04What we have to offer

Cash flows on day one, then tooling you could not have built from the same footage.

The commercial structure is deliberately simple: a fixed rate for every hour of accepted task data, per worker, multiplied across everyone enrolled. No cost basis to recover, so the margin is yours.

Unit economics

$

Rate per verified hour

A fixed rate for every hour of accepted task data, per worker — negotiated with the lab, then with you.

×

Hours × workers enrolled

Multiplied across every worker capturing data on the job. Enrollment is the only lever you control.

=

Cash flows to your business

A new revenue stream with no cost basis. Cash comes in as soon as the contract begins.

How the contract gets there

  1. 01

    Shop annotated pilot footage to all AI labs

    A short pilot produces annotated footage we take to the entire buy side, not a single counterparty.

  2. 02

    Choose the best contract

    We run the negotiation and select on rate, volume commitment, and term — then hold the commercial relationship.

  3. 03

    Bring the contract to you

    The deal comes back as a negotiated rate per verified hour, structured as a revenue share on accepted data.

Revenue model — illustrative

Interactive
250
12 hrs
$18
50%

Verified hours / month

12,990

Contract value / month

$233,820

Your royalty / month

$116,910

Your royalty / year

$1.40M

Illustrative only. Rates are set per contract and depend on country of origin, task complexity, and dataset diversity — the three drivers of dataset value. Real pricing comes back with your pilot scope.

01

A new revenue stream

A royalty on every accepted hour of data, with zero upfront cost and no disruption to your crews. Royalty income can offset labor, R&D, and other operating costs.

02

Efficiency & training dashboards

The same footage powers internal tooling: digital twins of your operations, surfaced safety and compliance risks, and training curricula built from how your best people actually work.

03

First in line on robotics

A direct line to frontier labs puts maturing robots and automation on your floor first. Your data pipeline becomes your automation pipeline — and you keep all the option value.

Partner dashboards

What you get back, beyond the royalty.

Live, interactive versions of these ship with the partnership. Frames below are structural placeholders.

01

Efficiency & Security Command Center

3D digital twins of your operations — optimize layouts, surface risks, run what-if scenarios without touching production.

Interactive · coming soon
OverviewFloorplanOperationsSafety & securityWhat-if simulator

Production efficiency

Throughput

OEE

On-time delivery

Safety incidents

Equipment uptime

02

Learning & Development Hub

Your footage, turned into role-based curricula and microlearning modules with your L&D team.

Interactive · coming soon
OverviewCurriculumLibraryLearnersAnalyticsReports

Training modules

Active learners

Completion rate

Knowledge retention

Skill proficiency

Safety incidents

03

First In Line Deployment Platform

Track robots maturing from frontier labs onto your floor first, alongside the annotated data lakehouse feeding them.

Interactive · coming soon
OverviewRobotics pipelineData lakehouseDeploymentsPerformance

Concepts

In progress

Piloting

Validation

Production ready

Revenue earned

05Examples

Where collection already runs — and what comes out of it.

We are servicing frontier labs across warehouse, residential, agricultural, and industrial sites today. Utility and industrial field services are next, and wide open. Open any environment for the pilot detail.

Next

Utility and industrial field services

Wide open. No meaningful dataset exists for this work today.

Discuss a pilot