The robot demonstration data your rig can't collect.
Bespoke, action-labeled, contact-rich human capture — delivered as RLDS / HDF5 / LeRobot trajectories to your ingest spec. We capture the environmental diversity and hand-object interaction your internal rig can't reach, and clear the rights so you can train a commercial model on it.
Compute is abundant. Action-paired, rights-clean data isn't — and it can't be scraped.
A funded market has formed on exactly this thesis — but the free datasets are your pretraining base, not the apex. What moves your eval metric is diversity, force, and legal cleanliness that a single internal rig physically can't produce.
You receive trajectories, not splats.
A policy has to output actions video doesn't contain. So we don't ship a 60-camera reconstruction — we ship the calibrated, action-labeled episode your pipeline already ingests: RLDS, HDF5 or LeRobot, to your exact schema.
episode/
├── observation/
│ ├── image.cam_high # (T, 720, 1280, 3) multi-view RGB
│ ├── image.cam_wrist # (T, 480, 640, 3)
│ ├── depth.cam_high # (T, 720, 1280) optional RGB-D
│ └── state # (T, 8) joint pos + gripper (proprioception)
├── action # (T, 7) 6-DoF Δpose + gripper ← what you train
├── hand_pose # (T, 21, 3) per-frame 3D finger joints (dexterous)
├── force_torque # (T, 6) contact-rich option
├── language_instruction # "pick the mug and place it on the rack"
└── meta/
├── calibration # intrinsics + extrinsics per camera
├── consent_id # per-contributor consent artifact ref
└── sync # genlock-validated timestampsYour rig captures three kitchens. Robots meet thousands.
Past a per-scene threshold, generalization scales with the diversity of environments and objects as a power law — far more than with raw demonstration count (ICLR 2025, arXiv:2410.18647). We supply the environmental, geographic and demographic spread your internal fleet physically can't reach.
Video alone can't train a policy. We capture the action.
Gyroscope, force/torque, joint encoders, hand pose — essential for some embodiments, noise for others. Here's exactly what each robot type needs.
| Robot type / task | Essential modalities | Nice-to-have | Video-only role |
|---|---|---|---|
Manipulation (VLA, parallel gripper) | Multi-view RGB + proprioception + 7-DoF action + language + calibration | Depth, wrist cam | Pretraining vision backbones only |
Dexterous / 5-finger manipulation | + per-frame 3D hand / finger pose (the recovered action) | Tactile | — |
Contact-rich assembly | + 6-axis force / torque + tactile (shear, slip, compression) | — | — |
Legged locomotion | IMU + joint encoders (proprioception core) | Vision / LiDAR for terrain | Minimal |
Drone / UAV | IMU + depth / RGB + state estimation → control commands | LiDAR, GPS | Minimal |
AV / ADAS not our focus | LiDAR + camera + radar + GPS/IMU + CAN-bus, time-synced | — | — |
Capture rig matched to the task — not one studio for everything.
Wearable multi-sensor
Lightweight human capture — head-mounted egocentric camera, motion-tracked gloves, IMU, optional fingertip force. The fast path to action-labeled human demonstration.
Volumetric multi-camera
A 60+ camera ring when occlusion-robust, view-invariant 3D hand + object pose matters. We use it to extract precise action tracks — the rig is the means, never the deliverable.
Genlock & calibration
Time-sync and calibration decide whether a multimodal dataset is trainable at all. Poor sync silently corrupts every observation–action pair. We validate it as an acceptance gate.
Legally safe to train a commercial model on.
Every contributor consented; every hour is commercial-use licensed; each buyer gets a fitness review. It's the one thing synthetic data structurally can't produce and offshore farms won't cleanly deliver. As an EU operator, we build the consent machinery as a first-class deliverable — GDPR as a feature, not a tax.
Matched to your embodiment — not over-specified.
Humanoid / whole-body VLA
Environmental & demographic diversity a single lab rig can't produce.
Parallel-gripper manipulation
The mainstream embodiment — matched, not over-specified with 25-joint tracking.
Dexterous 5-finger hand
Precise recovered hand action, occlusion-robust across viewpoints.
Contact-rich assembly
The friction, shear and compression physics simulation breaks on.
Legged locomotion
Proprioception-dense trajectories for blind and perceptive policies.
Pick the pilot closest to your need.
Early-access pilot pricing, subject to scoping — anchored above the ~$136/hr commodity teleop floor. No payment is taken; selecting a tier scopes our first call.
Probe
Pilot cohortSeed / pre-Series-A team validating whether bespoke capture moves their eval metric before committing budget.
- 1 task family across 1–2 environments
- Per-episode success/failure + language labels
- Single-buyer commercial license
- Delivered in RLDS or LeRobot, with calibration + a documented schema
- 24–72h scoping call; fast human follow-up
Cohort
RecommendedSeries A/B robotics or humanoid team scaling one VLA policy that has hit a diversity or rights wall.
- 3–5 environments with deliberate diversity (lighting, clutter, geography)
- Per-frame hand/joint pose where dexterous, contact timestamps
- Documented per-contributor consent artifacts
- Single-buyer commercial license (exclusivity optional)
- RLDS / HDF5 / LeRobot to your ingest spec
Scale
Diversity engineWell-funded humanoid / foundation-model team covering diversity gaps their in-house fleet physically can't reach.
- 8–15 environments; long-tail failure / recovery cases included
- Full multimodal stack incl. force / tactile on contact-rich tasks
- Quarterly refresh option
- Per-buyer fitness review + consent / model-release documentation
- Exclusivity option; delivery to your exact RLDS / LeRobot schema
Custom / Exclusive
By requestBuyer needing exclusivity, IP assignment, a recurring capture program, or a co-developed rig for a rare environment or embodiment.
- Data ownership / IP assignment terms negotiated
- Exclusive (non-resale) capture
- Recurring program option (capacity-subject, not a standing-inventory claim)
- Co-developed rig or new-environment buildout
- White-glove consent, compliance and acceptance program
Package price bands are early-access pilot estimates, subject to scoping. Any volumetric premium is exactly what this pilot exists to validate — it is not presented as an established market rate.
Specificity over logos.
We're a founding cohort, pre-launch, and we say so. No wall of customer logos — instead a real capture capability, an accountable EU operator, and a reference episode you can inspect under NDA before you commit a euro.
Real capture capability
Volumetric multi-camera rigs, wearable multi-sensor human capture, and 4DGS extraction — a concrete stack we own and operate, named on the capture page.
Accountable EU operator
Built and run by Flyreel Studio (EU · Poland). You contract with a real entity and talk to the person who owns the rig — not a reseller. DPA available on request.
Founding cohort, capped at 12
We onboard a small founding cohort so we can co-design each capture protocol. The cap reflects capture capacity — quality over volume, and exclusivity options for early partners.
Honest by construction
Every stat cites a public source. We claim a capability and bespoke collection — never a standing inventory we don't have, or logos we haven't earned.
The questions a Head of Data actually asks.
Why pay when huge open datasets are free?
Open data — Open X-Embodiment (1M+ trajectories), DROID (76k), EgoDex (829h) — is your pretraining base, and we say so. We sell the scarce apex it lacks: your environments, rights-clean human subjects, force-labeled contact tasks, and the failure cases open sets don't cover. Past a per-scene threshold, diversity (not volume) is what moves your eval metric — a power law, per ICLR 2025 (arXiv:2410.18647).
Can you deliver data our pipeline can train on? A 60-camera splat isn't trainable.
Correct — and we never ship splats as the deliverable. You receive RLDS / HDF5 / LeRobot trajectories: multi-view RGB(-D) + proprioception + a 7-DoF action label + per-frame hand/joint pose + language + calibration. The volumetric rig and 4DGS are the means we use to extract precise, calibrated 3D action and hand-pose tracks. There's a full sample schema on our data-format page — inspect it before you talk to us.
We collect in-house — why outsource?
Your rig captures a narrow environmental distribution — the field's most-cited weakness (a dataset from 3 university kitchens doesn't capture thousands of real ones). We supply the environmental, geographic and demographic diversity you physically can't self-produce, plus the long-tail failure cases you don't want to stage yourself.
Isn't sim / synthetic cheaper?
For the middle of the data pyramid, yes — use it. But simulators can't model the friction of a damp soap bar, foam give, or specular metal, and they can't give you real human dexterity or rights-clean human subjects. Buy real capture for the apex sim breaks on; synthesize the rest.
An EU boutique can't match offshore price or volume.
We concede price openly — we won't out-scale offshore teleop farms or the BPO giants, and we don't try. We win on modality rarity (force/tactile, dexterous hand pose), on sync/calibration integrity, and decisively on rights-cleanliness: consented, commercially-licensed human data an offshore farm can't legally guarantee for training a commercial model.
Who owns the data? Is it exclusive? Is consent airtight?
These are tier features, not footnotes. Default is a single-buyer commercial license with documented per-contributor consent artifacts and per-buyer fitness review; exclusivity (non-resale) and full IP assignment are available on the Scale and Custom tiers. As an EU operator capturing identifiable humans, we build the consent and model-release machinery as a first-class deliverable.
Scope a bespoke capture.
Four quick steps, mostly taps. Selecting a package scopes our first call — no payment is taken. Founding cohort capped at 12.
no payment now · reply within 1 business day · GDPR-compliant