Raw, real, physical demonstration data — captured on our own multi-camera rig, not simulated. Purpose-built for Vision-Language-Action (VLA) models and general-purpose robotics frameworks, with native export to RLDS and Hugging Face LeRobot.
Foundation models for physical AI need thousands of clean, diverse demonstrations of the same skill — not footage of one kitchen, one knife, one pair of hands. Most available datasets are narrow, licensed for research only, or missing from certain environments entirely.
And unlike synthetic or simulated alternatives, every trajectory in our data comes from a real body moving through real physics — 100% physical capture, 0% synthetic.
Purpose-built capture hardware and a two-node processing pipeline — not a generic webcam setup — deliver 3D coordinate matrices with no commercial IP constraints attached.
A high-throughput central node handles batch feature extraction and multi-view alignment. 2D joint detection (ViTransformer) is triangulated via DLT into direct 3D coordinate matrices — no proprietary mesh licenses, full commercial compliance.
| CAMERA RIG | 5× synchronized RGB — 1 top-view, 2 side-views, 2 wrist-mounted Global Shutter (AR0234) |
| KINEMATICS | Glove-based 6-DoF wrist & hand joint tracking, continuous grasp signal (0.0–1.0) |
| PIPELINE | ViTransformer → DLT / spatial triangulation → 3D coordinate matrices — zero commercial IP constraints |
| OUTPUT | Native HDF5 (.h5) and Zarr, structured for RLDS, Open‑X Embodiment and Hugging Face LeRobot |
| CONSENT | Written, scope-specific, commercial use disclosed up front |
Beyond our own dataset line, we capture bespoke data to your technical specification — new skills, new tools, new environments — using the same rig, pipeline and QC standard as EU-EgoKitchen Kinematics.
Trained for what comes next: Diffusion Policy and end-to-end manipulation policies for home environments, not just classification benchmarks.
EU-EgoKitchen Kinematics v1 is the proof: it's built on exactly this service, and it's the template we reuse for every custom order.
Our first dataset product: a kinematics-focused capture of fine-motor food-preparation skill in real European home kitchens. Delivered as raw HDF5 vector data — 3D joint trajectories, angular velocities and object poses — extracted through multi-view triangulation and hand-mesh recovery, not motion estimated by eye.
Tell us what's missing — embodiment, skill, region — and we'll tell you whether we can capture it.
