How much data a vision-language-action run needs, which container it should arrive in, and how to turn both into a capture brief.
A capture specification pairs first-person video with the telemetry a policy needs — IMU streams and a 3D camera trajectory — and every clip is annotated against the open FH-Ego v1 schema. Nothing here is stock: the brief defines the task, the environment and the rig before anything is recorded.
Every clip carries a SHA-256 consent receipt in a hash-chained, DPDP-aligned ledger, so a buyer re-verifies one clip instead of trusting a badge. Automated face redaction (YuNet) runs server-side at ingestion, before any human review, and fails closed — it refuses to emit a clip rather than emit one un-redacted. Licence plates are not covered. We hold no third-party attestation — no SOC 2, no ISO 27001.
Scoping is a conversation, not a price list: environment, session length, sensor rig and annotation tier each move the number. Capture Partners are paid transparently at a published band — ₹250–400/hr of accepted capture, by UPI on QA approval, no agent in the middle and no deduction. Those are the terms partners will be paid on, not a payout rail already running: the Capture Partner network is in pilot build-out and no payment provider is integrated.
Vision-language-action models such as RT-2 and OpenVLA emit low-level control from a language prompt and a camera view, so they need demonstrations that map vision straight onto coordinated action vectors. The figures below are published benchmark guidance for pretraining general VLA models — the field's figures, not our inventory. Read them as what to budget for, then scope a brief against the gap.
| Stage | Published guidance | What it buys |
|---|---|---|
| Imitation-learning bootstrap | 10k–50k demonstration clips | Basic task coordination |
| Robust policy generalisation | 200k–500k multimodal clips | Coverage of lighting, background and object variation |
| Held-out benchmark | 1k–5k calibrated clips | Reference trajectories to evaluate against |
Containers. We deliver LeRobot v2.1, HDF5
(EgoDex-style paired .hdf5 + .mp4, for ALOHA- and RoboMimic-style frameworks)
and the open FH-Ego v1 JSON schema; Zarr v2 and TFRecord (RLDS-style) are produced on request, and the
RLDS export is not tfds.load-compatible without a dataset builder — we would rather say so
up front. MCAP and WebDataset are not on offer: MCAP is a ROS 2 logging
container rather than a training format, and labs shard WebDataset internally.
Tell us the task, the environment and the policy you are training, and we scope a capture brief against it. Stated plainly: we are pre-revenue with no customers, one validated end-to-end annotation run, and a Capture Partner network in pilot build-out. The catalogue lists capture specifications you can commission, with target volumes and their real availability status.
The figures above are the published literature, not our inventory. If you want to know what a run of your size would actually take, that is a conversation rather than a table.