02What you receive

A LeRobot-format episode with a dataset card.

Per-frame columns for contact force and the sEMG feature vector behind it, gaze target, hand and head pose and demonstrator state beside the redacted video, in the format your loaders already read, with a card that prints its numbers or names the gate by which each is due.
Nothing offered for sale (ADR-0136)
01One episode, one clock

What every card prints.

Procedural illustration of one episode’s frames and columns on one clock; not a recording.

  1. sync_error_us

    Sync error

    p50, p95 and max in microseconds, measured from the LED-flash residuals on every recording.

  2. participant_count

    Participant count

    with the episode and frame counts; pseudonyms only.

  3. held_out_user_force_error

    Held-out-user force error

    the cross-user (LOSO) number against a vision baseline, or not run with the gate by which it is due.

  4. discard_counts

    Discard counts

    by reason: sync skew, missing consent, redaction failed, low quality, calibration invalid.

  5. redaction

    What was redacted

    the policy, the redactor's code hash, faces, bystanders, audio and frames dropped.

02The episode layout

The columns the writer lays down.

One parquet row per frame. A column is written only when the episode carries its stream, so a collection without a band has no force column and says so on its card. The vocabulary is the writer’s own (the LeRobot v2.1 writer in the data platform); the test suite pins this table to it. The video keys (observation.images.head, scene, or exo when the fixed third-person camera is on the session) are byte-for-byte copies of the redactor’s attested objects (monochrome today: faces and certified bystanders blurred, audio never captured).

ColumnTypeStreamWhat it isEgress
observation.hand_posefloat32[126]hand_poseTwo hands, 21 landmarks each, x/y/z in metres in the head-camera frame, estimated on the device from the camera frame before redaction; only the coordinates leave, beside the redacted video.Redacted with consent (beside the redacted video)
observation.hand_pose_confidencefloat32[2]hand_poseThe estimator's confidence per hand (left, right).Redacted with consent (beside the redacted video)
observation.head_posefloat32[7]head_poseHead position (x, y, z) and orientation quaternion (qw, qx, qy, qz).Redacted with consent (beside the redacted video)
observation.gaze_targetfloat32[3]gaze_targetWhere the wearer looked, on the video clock: a point in the scene-camera frame (u, v normalised, or x, y, z in metres only with a named depth source), NaN-padded to three values. Never an eye image; a gaze ray is not a shipped field.Features only
observation.gaze_target_classstringgaze_targetThe class of the object the gaze landed on, when labelled.Features only
observation.contact_forcefloat32[5]contact_forcePer-finger contact force in newtons (thumb, index, middle, ring, little), a calibrated estimate from the sEMG and IMU of the sEMG/IMU band after the 15-minute calibration; its held-out-user error is printed on the card.Features only
observation.contact_force_confidencefloat32contact_forceThe force model's confidence for the frame.Features only
observation.semg_featuresfloat32semg_featuresWindowed sEMG features with their schema version; never a raw sample window. Length set by the feature schema.Features only
observation.demonstrator_statestringdemonstrator_stateThe wire enum of fatigue and safety-relevant physical states, one value per frame, under a label-validity spec; never a mental state.Features only

Every episode also carries

  • timestamp
  • frame_index
  • episode_index
  • index
  • task_index
  • t_mono_us
  • t_utc_us
  • quality.q_score
  • sync_error_us
  • consent_id
  • redaction_policy_ver
  • cal_id

— the two clocks, the frame indices, the per-frame sync error, the consent id, the redaction policy and the calibration id behind the force estimate.

Beside the parquet, one video per camera

  • observation.images.head
  • observation.images.scene
  • observation.images.exo
03The dataset card

The card the exporter writes,
as it writes it.

This is the card make demo produced from a simulated recording, copied into the site unchanged and checked against the card schema. It shows the shape of every block: a measured number where the tool measured one, and not run with a dated gate where it did not. A collection on real sensors prints the same fields.

meta/dataset_card.jsonschema_version 4Synthetic example · source: simulation

cr4dle-demo-b030d2ed16f8

Written by cr4dle-dataset-export 0.0.1 on 2026-10-18T06:42:37Z (card schema version 4) from a simulated recording, by make demo. Not a collection; never release-eligible. The field paths in the margin are the card’s own keys.

Format
lerobot v2.1 · 30 fps · video codec none
Action space
finger_keypoints_v1 v1
Study
CR4-DEMO-2026-10
sync_error_us
Sync error, microsecondsp50 5000 · p95 5000 · max 5000 — measured from the LED-flash residuals of the simulated streams; the simulator’s detector floor.
participant_count · episode_count · frame_count
Participants, episodes, frames1 participant (a synthetic pseudonym) · 1 episode · 240 frames
held_out_user_force_error
Held-out-user force error: not runno LOSO regression report supplied to the exporter; due at the gate of . The block names its gate, so the card cannot leave the number out.
force_aware_vs_motion_only_ablation
Force-aware versus motion-only ablation: not runno ablation report supplied to the exporter; due at the gate of . The block names its gate, so the card cannot leave the number out.
sim_twin_correlation
Real-to-sim twin correlation: not runno real-to-sim correlation report supplied to the exporter; due at the gate of . The block names its gate, so the card cannot leave the number out.
identity_posture
Identity posture (re-identification audit)rank-1 62.0% against chance 20.0%, EER 18.0%, 5 participants × 2 sessions on observation.hand_pose, observation.head_pose.evidence_kind: synthetic · protocol reid_audit@2 · no eye image egress is a redaction fact, not an anonymity claim. Printed beside chance and never gated; a synthetic report from the benchmark fixtures.
discard_counts
Discard counts, by reasonsync_skew 0 · missing_consent 0 · redaction_failed 0 · low_quality 0 · calibration_invalid 0
redaction
What was redactedpolicy drop-all-frames-v1 · redactor sha256:84ba19e6abb7b4a65a8861e63415c114 · faces false · bystanders false · audio dropped true · frames dropped 240 · representation no_video (not_applicable). The simulated recording carried no video, so the redactor dropped every frame rather than blur a face it could not find.
consent_lineage
Consent lineage
  • consent-synthetic-45486bbf-25b1-5b04-8589-b7ae5b898127 @ synthetic-1 · purposes PURPOSE_ROBOT_TRAINING_DATA · granted 2026-09-13T00:00:00.000000Z · jurisdiction ZZ · scheme none · notice none · export scopes secondary_research, external_sharing, robot_training
A synthetic consent record minted by the simulator (jurisdiction ZZ). A real collection carries the scheme and notice version of its participant notice; consent is checked at capture and again at export.
stream_kinds · provenance
Streams and provenancestreams hand_pose, head_pose, gaze_target, demonstrator_state · devices dev-sim-glasses-0001 · firmware simulated · tiers TIER_L0_GLASSES · models demonstrator-state-stub-0, sim-hand-pose-0;metres-camera-frame, sim-head-pose-0, sim-gaze-0 · calibrations none · recordings rec-a740349ccfcd6b43, rec-313e607d1087113a · equipment none
episode_audit
Episode audithand-pose confidence lost below 0.5: mean 0% · p95 0% · max 0% over 1 episode with hand pose · regimes by_hand 1, leader_follower 0, vr_teleop 0, exoskeleton 0, walking 0 · gaze present 1, labelled 1, control-monitoring gaze excluded 0, scripted gaze excluded 0 · bystanders max 0 in 0 episodes
format.features
The feature table this export laid down
  • timestamp float32[1]
  • frame_index int64[1]
  • episode_index int64[1]
  • index int64[1]
  • task_index int64[1]
  • t_mono_us int64[1]
  • t_utc_us int64[1]
  • observation.hand_pose float32[126]
  • observation.hand_pose_confidence float32[2]
  • observation.head_pose float32[7]
  • observation.gaze_target float32[3]
  • observation.gaze_target_class string[1]
  • observation.demonstrator_state string[1]
  • quality.q_score float32[1]
  • sync_error_us int64[1]
  • consent_id string[1]
  • redaction_policy_ver string[1]
  • cal_id string[1]
  • action.finger_keypoints_v1 float32[133]
  • action.finger_keypoints_v1.valid bool[3]
A simulated Lens recording carries hand pose, head pose, gaze target and demonstrator state; a band in the kit adds the contact-force and sEMG-feature columns. The action-space adapter added action.finger_keypoints_v1.
04Formats

Two formats written.
Two named as plans.

A format is implemented when a writer exists and its layout is exercised by tests; a planned row is a plan, never a claim of support, and the exporter refuses it by name. This table mirrors the data platform’s format registry and is pinned to it by a test.

05The licence shape

Purposes no broader than consent.
A revocation duty on every licence.

The outbound licence register and its pilot template are in the repository as data with gates, before any paid team. ADR-0136 suspends the paid gates of this phase. The research gates stay in force. A licence names the dataset, the release and the object hashes; the register never holds a price or a name.

What is licensed
One released dataset at one release: exactly the objects whose hashes are in the delivery record, with the dataset card whose hash is named. The card is the description of the data; nothing not on the card is promised.
Permitted purposes
Named in the wire vocabulary of the consent record and never broader than the purposes the participants consented to, as printed on the card's consent lineage. A use outside them is outside the licence.
No re-identification, no onward transfer
The licensee does not attempt to identify a participant or bystander and does not pass the data on. Models trained on the data may be used and deployed within the permitted purposes.
Revocation duty
A participant may revoke at any time. On notice the licensee deletes the affected episodes from working copies within 30 days and excludes them from every later training run; no claim is made that an already-trained model forgets them. These are the register's defaults (ADR-0078), pending the founder's contractual decision.
What the register holds
Identities, dates, the revocation duty as three explicit values and hashes of the signed instrument, the card and the delivered objects. Not a price, not the name of a licensee: those live with counsel, outside the repository.
Price
Not announced. Cr4dle is a non-commercial research programme now (ADR-0136), so no sale occurs; no list price, plan or subscription exists, and no sale starts before a new decision replaces ADR-0136.
06The consent boundary

Features leave.
Raw signal stays.

Raw neural signal, sEMG and gaze never leave the device; they cross only as features. Video and the poses derived from it leave only after on-device redaction and under current consent, because every consuming pipeline trains on them. Consent is checked at capture and again at the instant of export, with revocation that is forward-looking; a revoked episode is not exported, and the discard is counted on the card.

Illustration of the device boundary: raw windows stay behind it; only feature vectors, pose coordinates and redacted frames cross. Not data.

Stays on the device

  • Raw sEMG and the sample windows behind the force estimate
  • Eye images and the raw gaze signal
  • Unredacted frames, faces, bystanders, audio
  • Raw neural signal from the research tiers

Crosses, with a card

  • Per-finger contact force and its confidence, as a calibrated estimate, and the windowed sEMG feature vector behind it (never a sample window)
  • The gaze target: a point in the scene-camera frame, on the video clock
  • Hand pose estimated on the device from the camera frame before redaction, and head pose (planned) from the camera and the IMU; only the coordinates leave, beside the redacted video
  • Redacted video (monochrome today: faces and certified bystanders blurred, audio never captured) under consent that is checked at capture and again at export, with revocation

Redaction is a mitigation, not an anonymity guarantee: the card’s identity posture prints a re-identification rate beside chance so the counsel of a dataset user can read it. No channel claims a mental state; demonstrator state names fatigue and safety-relevant physical states only.