Inside the
evaluation dataset.
Video, action labels and source-frame mappings in one inspectable delivery. Explore the fields, review methods and supported uses of our current V6 sample.
2 pages · English · Public sample brief · Updated 26 October 2026
| Field | Sample specification | Interpretation |
|---|---|---|
| Activity | 3 excerpts / 52 seconds | 30 s weighing and filling; 10 s sealing; 12 s floor retrieval. Independent selections. |
| Camera views | 6 per excerpt | Head, left wrist, right wrist and three external views. |
| Viewing video | 1920 × 1080 / 30 fps | H.264 viewing exports. Clip-relative playback timeline. |
| Action annotations | 16 dominant intervals | Editorial phase boundaries across the three selected operations. |
| Hand annotations | 21 joint slots per hand | 2D image-space estimates on the head view. Coverage varies; uncertain coordinates are null. Not six-view keypoint annotation. |
| Alignment | Relative QR-derived offsets | Relative scene alignment; no certified synchronization tolerance in this sample. |
| Traceability | 9,360 source-frame mappings | Mappings cover the six views across 1,560 selected time steps. File hashes support transfer-integrity checks. |
| Quality review | Documented sample-based checks | Detailed coverage and review records available for technical evaluation. No independent accuracy benchmark is claimed. |
| Not included | Metric 3D / robot controls | Current scope is human video and 2D annotations. Metric 3D, calibrated camera geometry and robot controls are outside this sample. |
Inside the evaluation delivery
The full evaluation package includes structured data and a local frame inspector. Request access to discuss a technical review; the website presents the face-blurred viewing version.
annotations/Action intervals, recovery events, hand estimates and schema definitions.annotations/frame_mapping.csvSelected frames linked to original source files and frames.hand_review.htmlOffline comparison of source images, model outputs and refined estimates.SHA256SUMS.txtChecksums for verifying delivered file integrity.
Planning a custom delivery
Define tasks, accepted duration, views, annotation depth and measurable acceptance criteria. A pilot checks whether the proposed collection meets your needs.
Delivery formats and integrations
Native delivery uses JSON and CSV. For LeRobot, RLDS, MCAP or other formats, share your target schema and reader requirements. Compatibility is evaluated through a separately scoped pilot; these adapters are not part of the current delivery.
3D geometry and hardware synchronization
These require equipment review, calibration and validation. Metric 3D poses and certified synchronization tolerances are not included in the current sample.
One environment.
Six perspectives.
Compare selected operations across the six recorded views. Alignment specifications are listed above.
What to know before integrating.
Can I directly train a robot policy with this sample?
This is a human-demonstration evaluation sample. It has no robot state or control-action trajectories. Suitability depends on your model, training method and required features; converting a file format does not supply missing supervision.
Are all hand joints annotated in every frame?
The schema provides 21 joint slots per hand. Available coordinates vary with visibility and estimation quality; unresolved coordinates remain null. Detailed coverage records are available for evaluation.
Does the recovery sample contain a recorded failure?
The clip shows retrieval from the floor and return to a tray. The initial drop and its cause are outside the recording excerpt; product condition is not established.
What are the licensing and access terms?
The public videos and PDF are for evaluation. They do not grant a dataset training, redistribution or resale license. Discuss the intended use and applicable rights with ULoveAI before a data delivery; any commercial license must be agreed separately.
Start with a technical evaluation.
Tell us your task, required fields and acceptance criteria.