Technical Field Note #02 · First-party capture evidence

Egocentric RGB + IMU Task Sample

A capability demonstration of Origin Data Lab's real-world capture pipeline, combining egocentric RGB video, native and application-level IMU streams, structured task metadata, integrity checks, and software-derived temporal alignment.

Published by Origin Data Lab · October 2026 · Technical evaluation sample

1920 × 1080 RGB video
30 fps 587 frames
410.641 Hz Measured native IMU
19.567 s Public video excerpt

Scope & Non-claims

What this sample demonstrates

This first-party technical sample documents one short task excerpt produced through Origin Data Lab's capture and verification workflow. It is intended to make measured capture characteristics, timing methodology, task structure, and delivery integrity directly inspectable.

It is a capability sample, not a fleet-wide performance claim and not an off-the-shelf training dataset. Temporal alignment in this release is software-derived. It does not claim hardware synchronization, frame-level synchronization, or sub-frame synchronization.

Capture Context

Real manipulation, structured as an inspectable task sequence

The excerpt contains food-preparation manipulation involving a knife, cutting board, onions, and a sweet potato. Approximate temporal task labels were created manually for this evaluation sample.

Public web preview · 1920 × 1080 · 30 fps · 587 frames · 19.567 s · no audio. The downloadable evaluation package retains the separately verified capture payload; this web copy is re-encoded only for browser delivery.
00.0–05.4 s

White onion trimming

05.4–07.0 s

Place / reach transition

07.0–15.8 s

Brown onion trim / peel

15.8–17.8 s

Task transition

17.8–19.55 s

Sweet potato trimming

Measured Sensor Characteristics

Measured rates and timestamp integrity

Stream Samples Measured rate Max interval Duplicate / non-monotonic
Native accelerometer 8,028 410.641 Hz 2.436 ms 0 / 0
Native gyroscope 8,028 410.641 Hz 2.436 ms 0 / 0
Application accelerometer 2,001 102.353 Hz 21.918 ms 0 / 0
Application gyroscope 1,998 102.162 Hz 21.918 ms 0 / 0

Native and application-level streams represent separate capture paths. In this sample, all 2,001 application accelerometer timestamps and all 1,998 application gyroscope timestamps are exact timestamp matches within the corresponding native streams.

The observed native-to-application accelerometer scale relationship is approximately 9.80665. This strongly supports application-level acceleration expressed in g and native acceleration expressed in m/s². Gyroscope values are consistent with rad/s, but implementation-level unit verification is outside the scope of this public sample.

Timing Methodology

Software-derived video–IMU temporal alignment

Sensor events carry nanosecond-resolution monotonic timestamps. For this public excerpt, the IMU streams were mapped to the public video timeline using capture-session timing anchors and visual–inertial refinement.

PRE-RELEASE CHECK A 0 ms R² = 0.798
PRE-RELEASE CHECK B +6 ms Pearson r = 0.929
STATED UNCERTAINTY ≤ 33 ms One 30 fps video frame

These checks were performed on the released excerpt and are therefore in-sample validation. They should not be interpreted as hardware synchronization or as proof of a session-invariant camera startup offset.

Delivery Structure

Inspectable files, task metadata, and integrity evidence

Capture payload

RGB video plus native and application-level accelerometer and gyroscope streams.

Structured task metadata

Task structure and interaction-event JSON provide approximate temporal descriptions of the demonstrated manipulation sequence.

Integrity & verification

SHA-256 checksums, an integrity report, and a verification script make package-level inspection reproducible. 

capture/
  video_preview_aligned_19p55s.mp4

sensors/
  app_level_accelerometer_clip.csv
  app_level_gyroscope_clip.csv
  native_accelerometer_clip.csv
  native_gyroscope_clip.csv

annotations/
  interaction_events.json
  task_structure.json

metadata/
  capture_metadata_public.json

quality/
  integrity_report.json
  verification_output.txt

tools/
  verify_odl_sample.py

checksums.sha256
DATA_CARD.md
LICENSE.md
README.md

Known Limitations

What this sample does not claim

The public package is a single short capability sample. It does not establish fleet-wide capture performance or production-scale acceptance rates.

Video–IMU alignment is software-derived and validated on the released excerpt. No hardware, frame-level, or sub-frame synchronization claim is made.

The sample contains manual approximate temporal annotations, not spatial hand, object, pose, segmentation, or 3D labels. Audio is not included in the public excerpt.

Camera intrinsics and extrinsics are not provided, and the physical mount configuration is not asserted in this release.

Technical Evidence

Inspect the released sample and verification tools

Access the public evaluation package on Hugging Face and inspect the verification toolkit, schema notes, and technical documentation on GitHub.

Hugging Face Evidence GitHub Evidence

Related Capabilities

From technical evidence to project-specific data production

This Field Note documents one inspectable capture sample. For project-specific collection programs, explore Origin Data Lab's related capabilities for egocentric human demonstration data, multimodal sensor capture, robotics training data, and real-world Physical AI data collection.

Egocentric Human Demonstration Data First-person task demonstrations captured in real working environments. Multimodal Sensor Data RGB, IMU, audio, metadata, and project-specific multimodal capture. Robotics Training Data Real-world task data structured around robotics learning requirements. Physical AI Data Collection Field production for Physical AI and embodied intelligence programs.

From Sample to Production

A closed-loop field data workflow

Origin Data Lab builds project-specific real-world data programs around the buyer's task definition, capture specification, site constraints, acceptance criteria, and required delivery schema.

Capture Ingest Validate Review Accept / Recapture / Reject Deliver

Evaluate Origin Data Lab

Need evidence against your own task and data specification?

Send us the task, environment, sensor, metadata, and acceptance requirements. We can scope an evaluation sample or paid pilot around your specification.