One capture.
Every dimension.
A real dishwashing capture, with IMU, LiDAR, camera poses, mesh, and MCAP, processed into scene understanding, hand motion, quality signals, privacy proposals, and training exports.
Recorded demonstration
Nothing is processed here.
Recorded results from a real 152-second capture. Playback uses a silent 720p web preview; overlay coordinates and annotations refer to the original 4K frames. Nothing is processed here.
Time accounting
Estimated activity across the whole video, weighted by source timestamps. Unknown time remains in the total.
| Visibility | Visible | Not both / out of frame | Unknown | Pose available |
|---|
For “Both hands,” false means at least one hand is not visible. Missing pose estimates do not establish that a hand is outside the image.
Action phases
Click to inspect the original moment3D hand pose
EstimatedRotate the view. Each hand has its own wrist origin; these are estimated poses, not calibrated world trajectories.
Privacy review
Download sidecar with saved redaction choices ↗Review candidates can be false positives. Preview blur does not alter the source video or bundled exports.
Measured quality
Which sections pass the current policy?
Candidate means the configured QC policy passed. It does not certify annotation accuracy, privacy clearance or robot-training readiness.
Inspect every QC interval
Thresholds and exact metric definitions
Training handoff
Real downloadable artifactsFrame-level JSON, research samples, and a LeRobot human-demonstration dataset. The official loader checks every exported observation. Human observations are not robot actuator commands.
Official loader validation is pending.
What a visual-learning sample looks like
A real frame paired with a later frame from the same continuous episode. These source previews are not a learned prediction or a privacy-cleared export.
Training integration check
A small network consumes the exported observations and performs real gradient updates. This checks the data handoff; it does not demonstrate robot skill or model accuracy.
Waiting for the exported training samples.
Post-training documents
The capture, in every dimension.
Loading fresh sensor audit…
The mesh and camera path were recorded. Depth points and hand-ray surfaces are alignment previews; spatial calibration is unverified. Hand-ray surfaces are visible depth samples, not anatomical joint centers. Grey path shows the full recording; amber shows elapsed movement.
Recorded LiDAR depth
Nearest recorded depth map · source timing
Exact recorded samples nearest this video moment
Depth fusion coverage and validity
Available after the fresh inference run completes.
Motion & sensor signals
Scrub the source video to move the white cursorAdditional motion: orientation, user acceleration, gravity, magnetometer and compass
Sampling rates, missing intervals and stream timing
| Stream | Samples | Measured rate | Longest gap | Non-increasing intervals |
|---|
Plots are reduced for display; the full source archive retains the original JSON and MCAP. Native IMU units were not declared by the capture, so acceleration and angular units are not guessed.
Every frame, inspectable
Aligned to original timestampsCoverage lanes appear after the fresh run.
Pose availability differs from hand visibility: an occluded hand may be visible without a reliable pose. Unknown periods stay in the denominator.
Hands, manipulation, holding, reaching, idle and unknown ; exact time accounting
Pending
Objects at the current frame
| Object / track proposal | Model score | Box · source pixels · xyxy |
|---|
Current frame: QC, usability decisions, action evidence and hand flags
Pending
Source integrity & synchronization
Download audit ↓MCAP recording contents
MCAP holds the time-stamped sensor streams. MAPLE checks and aligns them, then packages selected observations into training datasets such as LeRobot.
| Topic | Messages | Schema | Span |
|---|
Six fresh RGB synchronization checks
These images came from MCAP and were matched against the original MP4. Select one to inspect its matching video moment.
Calibration, coordinate frames and units
Annotation spot checks and observed limitations
Five illustrative action checks support the broad scene descriptions. Overlay checks also found missed bowl and plate boxes, missing poses for visible hands, and uncertain joints behind objects. These are model proposals; no numerical accuracy benchmark is claimed.
Inspect the specific findings ↓Open three overlay spot checks ↗Capture evidence & sensor exports
Enhanced annotations excludedInspect the original recording inventory, derived geometry, and audit evidence. Original source files are available on request. The annotation and LeRobot exports below belong to this recorded run.
Original capture files
Structured sensor outputs
Public downloads contain selected demo data only; full original sources are available on request. Privacy preview choices do not rewrite the bundled exports.