Diverse, license-clean human-demonstration data for contact-rich manipulation.
Egocentric video + IMU + hand-tracking of real skilled manual work, delivered as a LeRobotDataset. Per-field confidence and a MEASURED/INFERRED truth state on every label, cross-stream agreement as the anti-fabrication signal, and a datasheet per clip.
We resolve what we can first. Then we ask only what we can't.
The review session is not a survey — it is a reasoning-exhaustion funnel. The system infers, audits, and cross-checks everything it can before it ever asks the worker a question. Only genuinely worker-exclusive information escalates. This is the architecture; the published buyer-facing session populates from the live run once it lands.
action: "seat fastener"evidence_span_ms: [161240, 162480]truth_state: INFERRED · confidence: 0.83cross_stream: video↔wrist_imu AGREEtruth_state: WORKER_CONFIRMED — if answeredtruth_state: UNKNOWN — if not yet answeredNeither state inflates the otherArchitecture diagram — not a count of confirmed labels. Published buyer-facing sessions populate from real worker answers; no confirmed labels ship until a real answer lands.
The label, the evidence, and the review state — synced to the footage.
Our viewer plays the raw episode alongside the structured layers a model team actually audits: an evidence timeline synced to the video, an artifact-status panel that marks every surface missing / degraded / inferred / confirmed, and a signal-quality panel showing cross-stream agreement as visible anti-fabrication evidence.
Our pipeline processed the license-clean HO-Cap hand-object dataset (CC BY 4.0, attribution rendered in-viewer). Scrub the episode, inspect each layer, and verify the per-field truth state.
Open the sample episode →One action span. Two independent expert reads. One annotated record.
The same audited understanding of each action span carries two expert-grade dimensions: the worker's own account of intent and physical context, and an independent safety / code-compliance read against the codes a robot must meet to operate on a real regulated jobsite. These lanes never cross-contaminate — the audited-linkage panel shows which evidence backs which claim.
actionseat fastener with impact driverevidence_span_ms[161240, 162480]confidence0.83cross_streamvideo↔wrist_imu AGREEwhy_claimWorker-stated intent and situational context for this actionphysical_noteGrip variant, body position, load condition — worker-exclusive knowledgePending real Clerk-bound worker answer. No confirmed WHY label ships until it lands.
behavior_quality_labelObserved interaction against the codes a robot must meet on a real regulated jobsite — OSHA, NFPA, applicable trade codesseverity / code_refSeverity tier + specific code citation for the observationCredentialed safety role is currently the founder. PROFESSIONAL_* truth state reaches confirmed once a credentialed non-founder expert answers.
Architecture diagram. The diagram shows the mechanism and the honest pending state — not a population of dual-confirmed labels. Labels claim the architecture and what will publish; actual confirmed counts render in the live sample once real answers land.
Every field carries its truth state. Agreement is the confidence signal.
We don't flatten uncertainty into a single score. Video↔wrist-IMU agreement is computed per-moment and published as the confidence signal — it dips honestly during occlusion, which is precisely what monocular-only methods can't replicate. Broader multi-modal fusion (phone IMU, audio, scene-OCR) is the build direction; the wrist-IMU agreement layer is what ships today.
MEASUREDRead directly from a sensor or pixel — no inference.INFERREDDerived from evidence; carries its confidence score.WORKER_CONFIRMEDWorker reviewed + confirmed via the session.DEGRADEDPresent but low quality; signal below threshold.UNKNOWNNot determinable from available evidence.REJECTEDFailed a named gate — never shipped as fact.
Illustrative signal shape — not from a specific clip. Real agreement vectors render in the live sample. During full occlusion, the dominant-wrist IMU continues to produce MEASURED kinematics where monocular pose estimation produces nothing.
LeRobot v3 / RLDS in three lines. Croissant datasheet per clip.
Every clip ships as a LeRobotDataset (v3 Parquet layout, RLDS-compatible) with per-field confidence and truth-state as sibling columns — so your training loop can filter by epistemic quality without post-processing. A scored Croissant 1.1 datasheet ships with each clip, following the Gebru datasheet format with modality-presence matrix and named gates.
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("trades-data-ai/hocap-demo-01")
# Every observation column has a sibling truth-state column:
# obs.wrist_imu.gyro_rad_s → MEASURED (direct sensor read)
# obs.wrist_imu.gyro_rad_s.ts → truth state enum
# obs.video.contact_active → INFERRED (cross-stream agree)
# obs.video.contact_active.ts → truth state enum
# obs.why_claim → UNKNOWN (pending worker review)
#
# Filter to only worker-confirmed episodes:
confirmed = ds.filter(
lambda ep: ep["obs.action_truth_state"] == "WORKER_CONFIRMED"
)
# RLDS-compatible — works directly with Open-X loaders.Per-clip datasheet format — live scores render once the GPU run lands.
Your policy doesn't need more demos. It needs more environments.
The published evidence points the same way: diversity of environments and objects generalizes better than sheer demonstration count, teleoperation is throughput-capped, and pooling data across embodiments transfers. We present this as cited industry context — not our own results.
External research, cited for context. Figures are the authors' reported numbers, not Trades Data AI measurements.
Publish rate below 100%. Flagged episodes are data, not waste.
Every clip that fails a named gate produces a structured rejection — never a padded result. Flagged episodes carry the gate they failed and the reason, which is itself signal for your eval pipeline (failure-recovery, edge-case distribution). The quality heatmap and per-gate pass/fail render in the live sample; below is the gate taxonomy.
Measured accuracy vs. ground-truth belongs to the first non-founder (MMC pilot) worker run, gated on MANO counsel clearance for the hand-pose layer. Our single founder-recorded methodology clip is n=1 — an accuracy headline from it would be meaningless as a generalization claim. We will publish it honestly when it is real: first non-founder worker, real eval dataset, MMRV / RoboArena rank-correlation cited next to our own figure. No fabricated number ships first.
License-clean, consented, provenance-tracked, human-verified.
Most public ego/manipulation corpora are non-commercial, and acquisition method is now legally decisive. Our differentiator isn't volume — it's data you can actually deploy and defend.
License-clean & consented
Recorded by consenting workers under signed agreements — not scraped. The acquisition method is documented per clip, so you can deploy commercially without a provenance time bomb.
Provenance you can verify
Every published claim links back to the decision and code that produced it, with an in-browser hash recompute. Inspect the chain; don't take our word for it.
Datasheet per clip (Croissant)
A Gebru-style datasheet and Croissant 1.1 metadata ship with each clip, with LeRobot / RLDS load snippets — load it in three lines.
Human-verified truth state
Workers review and confirm labels; cross-stream (video↔IMU) agreement is the published confidence signal. MEASURED / INFERRED / WORKER_CONFIRMED on every field — never confidence theater.
One spine, two scope-tagged products
The same audited understanding emits the robotics dataset and a code-compliance safety read, with no cross-lane leak — the audited-linkage panel shows the lanes per decision.
No fabricated or synthetic-as-real data
Missing renders as missing; a failed gate is a structured rejection, never a padded result. We publish what we don't know.
What we sell. What we don't.
Expert buyers trust a vendor that discloses its limits. Here is what we don't sell — which, to this audience, is part of the credibility.
Evaluate a sample episode, then the datasheet.
We'll share a sample episode (RLDS / LeRobot-ready) and the per-clip datasheet so your team can judge usefulness before any commitment.