Services · Custom data collection

Custom robotics data collection

Custom robotics data collection is teleoperation or egocentric data captured to your specification — the tasks, robots, scenes and volume you need — rather than bought off the shelf.

PrismaX scopes each order against its own fleet and the VLA Foundry operator network, starts collecting within days, and delivers QA-scored episodes.

5,000 hrs
Collection capacity / month
33
YAM dual-arm rigs in our own fleet

What you can specify

Your tasks, your robots, your scenes.

Tell us the tasks, robots and scenes you need — you get exactly that.
  • Tasks. The tasks your model needs to learn.
  • Robots. The embodiment you train on. The catalog covers RLM 63, YAM, Piper, TOK2; Enterprise supports 12+ robot arm types.
  • Scenes. The scenes you need. The catalog spans home, industrial, kitchen, lab and research, logistics, office.
  • Volume. Up to 5,000 hrs of collection a month.

Process

Request → Capacity → Delivery

  1. 01

    Request

    Tell us the tasks, robots and scenes you need, and how much of it.
  2. 02

    Capacity

    We scope the order against our own fleet and the VLA Foundry operator network, quote it, and start collecting within days.
  3. 03

    Delivery

    Episodes are reviewed against the Standards for Data, QA-scored, and delivered by browser download, API or JSON export.

Our capabilities

Built to collect data that trains.

We run the whole loop ourselves: hardware, operators, review, and training on the result.
Fleet

33 YAM dual-arm rigs

Our own fleet of bimanual YAM rigs, so collection doesn't wait on someone else's hardware or schedule.
Network

VLA Foundry operators

Owner-operators and reviewers who capture and score data and get paid for it — $1M+ paid so far. How VLA Foundry works.
Egocentric

Real work sites

First-person capture of hands-on work at real work sites. See egocentric data.
Quality

Human validators

Every episode passes automated checks, then validators judge the demonstration itself: pass/fail criteria plus a sliding-scale score.
Reproducibility

Standardized rigs

Individually measured URDFs for every rig in the fleet, globally available camera SKUs, and standardized teleoperation code.
Training

We train on what we collect

An in-house team fine-tunes and evaluates models on our data and feeds the results back into collection. See our π0.5 fine-tuning notes.

Pricing and timeline

Priced below doing it yourself.

Custom collection is part of the Enterprise plan, which also includes the full data bank.
Pricing modelPer agreed contract
BenchmarkPriced against the cost of collecting it yourself
StartWithin days of scoping
CapacityUp to 5,000 hrs / month
DeliveryBrowser download, API, JSON export
Minimum orderTBD
Time to first deliveryTBD

What you get

Episodes like these

Teleoperation
Bimanual flower arranging, overhead and wrist cameras

Two robot arms arrange flowers in a vase, seen from the overhead and wrist cameras at the same time.

Quality review
Training-grade versus rejected episodes

Flower-arranging episodes from three cameras, each labeled training-grade or rejected after review.

FAQ

Frequently asked questions

What is custom robotics data collection?

It is robot training data captured to your specification — the tasks, robots and scenes you need, in the volume you need — instead of bought off the shelf. PrismaX collects teleoperation data on its own fleet and through the VLA Foundry operator network, and egocentric data at real work sites.

How fast can collection start?

Collection starts within days of scoping. Off-the-shelf datasets from the catalog can ship the same day if they already cover what you need.

How is teleoperation data collection priced?

Custom collection is priced per agreed contract, against what it would cost you to collect the data yourself. Capacity runs up to 5,000 hours of collection a month. Contact sales for a quote.

Can you collect on the robot we use?

Tell us the embodiment you train on. The catalog covers YAM, Piper, RLM 63 and TOK2 arms and Enterprise supports 12+ robot arm types; for anything else, describe the embodiment and we scope production around it with you.

How do you make sure the data is good?

Automated checks catch the obvious failures. Then human validators review each demonstration against the Standards for Data: pass/fail criteria such as full overhead view, clean lighting and a task that matches the prompt, plus a sliding-scale score for the quality of the demonstration itself.

Should I buy from the catalog or order custom collection?

Start with the catalog if its tasks, scenes and robot arms cover your training objective; Discover and Researcher give self-serve access. Order custom collection when you need a missing task, an unfamiliar environment or a specific embodiment.

Next step

Tell us what your model needs to learn.

Share the tasks, robots and scenes. We'll come back with scope, capacity and a quote.