not published
confidence ceiling not publishedBuyer dataset viewer
hocap-demo-01
Raw clip, action evidence, WHY chain, worker review, closure audit, and standards status.
Action timeline
Truth-state spans
Sensor evidence overlay
Audio and IMU evidence tied to player time
Signal-quality preview — what a buyer evaluates before signing
Clip stats, action-span profile, and cross-stream agreement (anti-fabrication evidence)
Action-span length distribution
- <1s0
- 1-3s11
- 3-10s0
- 10-30s0
- >30s0
Cross-stream timing agreement
Cross-stream alignment artifact is not published for this clip.
Selected span WHY detail
wait
- description
- worker sits at table with right hand resting on surface; no object interaction occurs during this interval
- truth state
- inferred candidate
- confidence
- 0.80
- granularity
- 5.00s
Evidence chain
WHY claims and citations
Worker review
Closure audit
Publish readiness
Audit depth — the verifiability moat
Verify it yourself: provenance, measured accuracy, signing
Every shipped decision carries a re-derivable provenance record, a worker-confirmed accuracy measurement, and an honest signing status — rendered exactly as produced. Missing surfaces render as missing, never mocked.
Per-decision provenance — the verifiability moat
Verify it yourself
Audited linkage — the per-decision spine
One spine, two products
Measured accuracy (worker-confirmed)
/lab-clips/measured-accuracy/hocap-demo-01.json
{
"accuracy": null,
"artifact_truth_state_counts": {
"action-timeline": {
"_empty": 1
},
"why-claims": {
"_empty": 1
}
},
"caveats": [
"n_worker_answers=0: no calibration data yet; scaffold only."
],
"clip_id": "hocap-demo-01",
"delta_truth_state_counts": {},
"n_distinct_workers": 0,
"n_worker_answer_deltas": 0,
"schema_version": "measured-accuracy-v1",
"status": "SCAFFOLD_NO_CALIBRATION_DATA",
"wilson_bounds": {
"estimate": null,
"lower": 0,
"upper": 1
}
}{
"action-timeline": {
"field_count_by_truth_state": {
"_empty": 1
},
"total_fields": 1
},
"why-claims": {
"field_count_by_truth_state": {
"_empty": 1
},
"total_fields": 1
}
}[ "Worker-confirmed label accuracy = fraction of worker_answer deltas where the auto-label was CONFIRMED UNCHANGED (WORKER_CONFIRMED). Corrections (WORKER_REJECTED) count against.", "Conformal calibration (MAPIE/crepes) requires a real calibration set; current bounds are Wilson interval prior — not calibrated.", "PROV blocks are W3C-PROV terms as dict keys (no rdflib). Render as RDF-PROV if a full graph is required." ]
C2PA / provenance signing status
Missing artifact
Raw integrity
Missing artifact
Semantic index
Missing artifact
Bundle status
Missing artifact
Modality matrix
Present, missing, or degraded streams
not published
confidence ceiling not publishednot published
confidence ceiling not publishednot published
confidence ceiling not publishednot published
confidence ceiling not publishednot published
confidence ceiling not publishedStandards scorecard
All dataset standards
Why this dataset — verifiability beats claims
Differentiators, honest limitations, and how to buy more
The differentiators (lead with how quality is enforced)
We ship rich human (21-keypoint + per-joint angles + wrist 6-DoF), you retarget to your hand — the field-standard interface (HumanEgo/EgoVLA/EgoScale), not pre-retargeted to one robot.
Contact is keypoint-in-named-mask cross-checked by IMU/audio — 2D overlap alone is wrong on near-misses; ours is sensor-corroborated, with the per-moment agreement shown in the signal-quality panel.
Every claim drills to its receipts and re-derives; the Merkle root recomputes in your own browser. Fine-grained queryable provenance is a documented unmet bar — it is the premium tier here.
The reasoner interrogates its own uncertain claims, asks the highest value-of-information question, and a worker's gated answer flips the field to WORKER_CONFIRMED (or holds it DEGRADED if unearned) — the elicited intent/rationale footage structurally cannot contain. Wired on the live path; the promotion gate refuses any synthetic/unbound identity, so a confirmation requires a real worker bound by the authenticated web layer — never faked.
Every field carries MEASURED/INFERRED/WORKER_CONFIRMED/DEGRADED/UNKNOWN/REJECTED + confidence + provenance. A model-card-grade limitations section is buyer risk-reduction, not weakness.
Fail-closed audio consent, PII-blur every frame, lawyer-reviewed contracts — your training-data documentation de-risking (EU AI Act), made a selling feature.
The manipulation tier (humanoid-core) and the project-execution tier (heterogeneous-team orchestration) — the second shipped INFERRED / north-star, honestly.
Honest limitations — stated, then refuted
- Object 6-DoF + near-field depth are DEGRADED (monocular, non-metric).The wrist IMU anchors global wrist pose independent of vision (survives occlusion), and every pose ships with per-moment cross-modal confidence — the differentiator monocular-only cohorts lack. We never label a monocular guess MEASURED.
- Corpus scale is early — few fully-populated clips today.Raw-data volume is not the moat; generalization follows a power law in environment/object DIVERSITY (Lin et al., ICLR 2025), and the scarce, un-replicable real-world distribution + expert-confirmed WHY is the defensible asset. We sell the data + audited annotations, never the model.
- Dense reward-progress is not in the export yet.It is EVAL-GATED: the TopReward layer ships on the LeRobot/RLDS export ONLY after it clears our eval (construction-egocentric transfer is unproven). Until then the channel is an honest placeholder — never a fabricated dense signal.
How we argue value
We report in the vocabulary you trust — real-world-success correlation + MMRV (SIMPLER / WorldEval lineage): “policies trained on our data rank / perform correctly,” not a bespoke internal score. Model-lift is reported per bundle where available — never asserted without the comparison run behind it.
Buy more — the round-trip
- Request a slice / label / trade — tracked requested → in-production → delivered.
- Subscription, not a file — changelog, version tags, refresh cadence; newer data is one click.
- Quality is contractual + visible — per-field truth-state + confidence + dated quality history; a failed gate shows REJECTED, never silently shipped.
- The corpus is flywheel fuel — each worker correction compounds into reusable, citable rules; honest scarcity is the moat.
Datasheet (Gebru et al.) + Croissant 1.1 + load-in-3-lines
Dataset card — credibility + machine-discoverable metadata
1 · Motivation
- purpose
- Robotics human-demonstration dataset + an insurance safety-observation feed — one scene understanding, two scope-tagged products.
- product surface
- not published
- intended use
- buyer evaluation, model fine-tuning, design-partner pilot, insurer loss-control review
2 · Composition
- modalities present
- see modality presence matrix
- feature columns
- not published
- proprio statistics
- not published
- truth-state distribution
- not published
3 · Collection process
- worker consent
- not published
- audio consent
- not published
- camera calibration
- not published
- reproducibility
- not published
4 · Preprocessing / cleaning / labeling
- PII redaction
- not published
- C2PA / provenance
- not published
- model versions
- not published
- git sha
- not published
5 · Uses
- format export ceiling
- not published
- not-intended use
- re-identification, worker discipline, redistribution beyond license, treating DEGRADED/UNKNOWN as ground truth
- cancelled / excluded surfaces
- world-model, dense mesh, digital twin, robot policy (per cancellation boundary)
6 · Distribution
- license
- not published — not published
- license posture
- not published
- serializations
- LeRobot v3 (canonical), RLDS / Open-X, sim-ready kinematic export
- license chain
- not published
7 · Maintenance
- codebase version
- not published
- generated at
- not published
- decision-precedent coverage
- not published
- closure audit
- not published
Two altitudes from one capture
Humanoid-core: per-finger articulation + wrist 6-DoF + cross-modal-verified contact + the action/observation representation a robotics buyer trains on. MEASURED/INFERRED per field.
The team-orchestration graph for heterogeneous robot teams (who does what, in what order, with what dependencies). Shipped INFERRED / north-star and labeled as such — never asserted MEASURED.
Load in 3 lines
LeRobot v3 (canonical)
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("trades-data/hocap-demo-01") # LeRobot v3 layout (in the buyer bundle)
print(ds[0]["observation.truth_state"], ds[0]["action.label"])RLDS / Open-X
import tensorflow_datasets as tfds
ds = tfds.load("trades_data_hocap_demo_01") # RLDS / Open-X episode-step
for step in next(iter(ds["train"])).numpy(): ... # same columns + truth_state passthroughCroissant 1.1 metadata header
License + modalities + size + task categories + consent flags as machine-actionable usage policies. This is the metadata header; the full RecordSet ships inside the LeRobot/RLDS export.
{
"@context": {
"@vocab": "https://schema.org/",
"cr": "http://mlcommons.org/croissant/",
"conformsTo": "dct:conformsTo",
"dct": "http://purl.org/dc/terms/"
},
"@type": "Dataset",
"conformsTo": "http://mlcommons.org/croissant/1.1",
"name": "trades-data-hocap-demo-01",
"description": "TRADES_DATA Tier-1 human-demonstration dataset (robotics) + safety-observation feed (insurance) from a single audited scene understanding. Every field carries a per-field truth-state + confidence + provenance.",
"license": "see dataset card",
"version": "unversioned",
"keywords": [
"robotics",
"human-demonstration",
"egocentric",
"construction-trades",
"manipulation"
],
"creator": {
"@type": "Organization",
"name": "TRADES_DATA"
},
"cr_size": {
"feature_columns": 0,
"proprio_statistic_columns": 0
},
"cr:usageInfo": {
"two_party_consent_audio": "see dataset card",
"pii_redaction": "see dataset card",
"worker_consent": "see dataset card",
"holdout": "clips flagged holdout in the DB never enter any buyer bundle (write-path guard)",
"not_for": [
"re_identification",
"worker_discipline",
"redistribution_beyond_license"
]
}
}