Skip to content
SyncRobotic
Capture & dataset engine

An egocentric robot-data factory, in your pocket

Record synchronized RGB, LiDAR depth, 6-DoF pose and IMU into one open MCAP file. VLMhub turns every recording into a queryable, training-ready dataset on its own.

recording.mcapros2 · cdr
rgb
15 Hz
depth
15 Hz
pose
60 Hz
imu
100 Hz

synced on header.stamp · depth ±0.1s · imu ±0.05s

One file, four synchronized streams

A proprietary on-device pipeline masters the MCAP contract, so the capture is compatible with the open stera-sdk with no cloud dependency. Every stream carries a correct header timestamp.

RGB

15 Hz

720p or 1080p JPEG, the master clock every stream syncs to.

LiDAR depth

15 Hz

256 x 192, 16-bit millimetre depth, registered to the RGB frame.

6-DoF pose

60 Hz

ARKit camera-in-world transform, captured every tracked frame.

IMU

100 Hz

Linear acceleration, angular velocity and orientation.

From raw capture to training-ready, untouched

  1. 1

    Capture

    The app writes one streaming recording.mcap on device, crash-safe against thermal-throttle frame drops.

  2. 2

    Atomic ingest

    VLMhub catalogs each recording in a single atomic transaction, so no failure leaves the catalog inconsistent.

  3. 3

    Enrich

    Hand-pose tracking is dispatched as a background worker, decoupled from ingest throughput.

  4. 4

    Quality gate

    Blur and tracking-confidence scoring decide accept, reject or retry, with no human in the loop.

  5. 5

    Export

    Timestamp-indexed seeking produces PyTorch-ready episodes for vision-language-action training.

An autonomous data-engine, not a viewer

VLMhub treats every new recording as a job to be driven to completion on its own. The result is a data-flywheel for embodied agents: zero human touch from raw capture to a queryable, training-ready dataset.

  • Decides the processing path: fast-ingest versus deferred enrichment
  • Dispatches ML hand-pose tracking as a background worker
  • Makes quality-gating decisions and retries on its own
  • Self-heals via crash-safe atomic transactions
Next in the loop

TuneKit

Once you have a training-ready dataset, TuneKit's self-driving agent turns it into a deployed, domain-specialized model.

Explore

Bring your robot-data pipeline in-house

Tell us about your capture program or private-model project. We will walk you through a self-hosted deployment end to end.

Get in touch
info@syncrobotic.space
Book a demo