Physical AI Training Data · Made in Europe

We teach
robots how
to greifen.

GDPR-compliant egocentric motion datasets from Germany and the EU — annotated with 21 hand keypoints in 3D, natural language descriptions, and delivered LeRobot-ready for Physical AI training.

21
Keypoints per hand
3D
Spatial depth (z_norm)
EU
GDPR + AI Act compliant

Datasets built for
Physical AI teams.

Every dataset is GDPR-documented, egocentric, annotated with 21 hand keypoints in 3D, and delivered in LeRobot format. Pricing is scoped per engagement — no standard shelf pricing.

01 — EVERYDAY MOTION

EU Everyday
Datasets

Egocentric hand and body motion from daily life across Germany and the EU. Diverse participants, 21 keypoints x/y/z, natural language frame descriptions. Ideal for general manipulation pre-training.

View specs →
HouseholdCraftsCareKitchen
02 — INDUSTRIAL

Industrial
Datasets

On-site recordings in real German warehouses and production lines. Workers performing real picking, assembly, and tool-use tasks. Failure cases included. Exclusive access only.

View specs →
LogisticsAssemblyFailure Cases
03 — EQUIPMENT AS A SERVICE

EaaS —
Your Facility

We install our recording infrastructure at your facility. Monthly datasets, your tasks, your workers, your environments — delivered robot-ready on a recurring basis.

Request a quote →
On-siteMonthlyCustom Tasks

Five steps to
robot-ready data.

01
Record
4K egocentric video in real environments. Consent-documented.
02
Transcode
FFmpeg → 1080p H.264. Frame extraction at 2–5 fps.
03
Auto-annotate
MediaPipe: 21 keypoints + z_norm. Grounding DINO for objects.
04
Human review
CVAT correction + natural language per frame.
05
Deliver
LeRobot · COCO 1.0 · YOLO · GDPR docs.

The data no US provider
can replicate.

🔒

GDPR by design

Explicit informed consent for every participant. Full DPIA documentation. From August 2026, EU AI Act enforcement requires GDPR-compliant training data for GPAI models — we've been compliant from day one.

🏭

Real industrial access

We deploy equipment on-site in German warehouses, production lines, and care facilities. Gig-workers recording at home cannot replicate factory-floor conditions. Physical access is the moat.

📐

EgoScale-validated approach

The EgoScale paper (arXiv Feb 2026) established empirically that egocentric data improves robot grasping success by 54%, with a near-perfect log-linear scaling law (R²=0.9983). Our pipeline is built around this finding.

🇪🇺

Independent from Big Tech

Following Meta's €14B investment in Scale AI, European robotics teams are actively seeking independent data suppliers. greifen is structurally independent — no Big Tech affiliation, no vendor lock-in.

Built for robotics teams
that ship.

Our datasets are used by teams training humanoid robots, manipulation models, and Physical AI systems across Europe and beyond.

Humanoid Robot Companies
Automotive OEMs
Logistics Operators
Research Institutes
AI Foundation Labs
Robotics Startups

Ready to train
better robots?

Tell us what you're building. We'll propose a dataset scoped to your task within 48 hours.

No commitment. We respond within 48 hours.