Synchronized sessions
Time-aligned sensor streams, calibration references, session manifests and processing status on a shared timeline.
Product / Actuate
Actuate turns heterogeneous real-world capture into synchronized, provenance-carrying episodes, quality reports and datasets that can be loaded by the customer's training stack.
What Actuate produces
Actuate makes heterogeneous capture inspectable, comparable and loadable without overstating what each sensor or stage measured.
Time-aligned sensor streams, calibration references, session manifests and processing status on a shared timeline.
A consistent, versioned representation of observations, task structure and actions across heterogeneous capture systems.
Per-component quality results, source lineage, consent and privacy status, and explicit reasons when a stage did not run.
Processing stages
Each stage records what ran, what did not and why. Supported recovery depends on the capture path and validation status.
Confirm the declared rig, channels, metadata and session integrity.
Align video, IMU, audio, pose, tactile and robot streams onto a shared timeline.
Recover hands, objects, depth, task intervals and action structure where supported.
Map heterogeneous capture systems into a consistent, versioned episode representation.
Record data quality, processing status, provenance, consent and privacy results.
Deliver LeRobot, RLDS, canonical or agreed customer schemas.
Real processed output
Scrub a real processed session in Rerun. The recording is repository-owned internal evaluation output, not a mock and not customer data.
The web recording is an optimized excerpt of a larger pipeline output. Video and depth display assets were reduced for serving; geometry, timeline, certificate values and annotations remain from the run.
Delivery record
Each episode states what ran, what passed and which export format was validated.
Monocular processing is validated on real capture. Other modalities retain their integration or embodiment-specific status.
Channel integrity, lineage, quality and privacy results remain attached to the versioned episode.
Repository acceptance covers LeRobot v3 and RLDS. Versioned and customer schemas are mapped per program.
Technical validation notes
Repository acceptance evidence and limitations are kept here without interrupting the product story.
The source pipeline recording was 520 MB. The web version downscales video to 960 × 540, depth display to 320 × 180 and decimates the point cloud. The session is internal evaluation capture; no customer data is shown.
Ingestion checks filename provenance, stream layout and stereo correspondence before assigning a rig contract. Stereo ingestion verification exists; calibrated stereo depth remains unimplemented pending end-to-end validation.
The acceptance evidence reports sync integrity 0.88 and calibration completeness capped at 0.30 because intrinsics were approximated. Unavailable measurements remain explicitly not measured. The displayed 3 / 5 result is normalized only over measured components.
The repository records 619 exported frames loaded through LeRobotDataset and 619 RLDS steps loaded through TFDS. These are pipeline acceptance checks, not model-performance claims.
Unknown consent or PII status blocks export. The recorded retargeting candidate failed kinematic checks and was withheld while human-space data remained available.