02The human sensorium · a proposal

Video shows the motion. People feel the contact.

An agent in a human body must learn how hard to press, when a foot takes weight and where the eyes go first. Video does not show these things. The capture kit is to measure them from people, one channel at a time. Each channel is a dated bet, and no channel is built or recorded.
01What an agent must learn

A body to operate.
Data to learn from.

Each skill below needs measured contact, gaze and event labels on one clock. Each label is to carry a measured sync error and a held-out-user label error.

What an agent in a human body must learn, and the data from people that teaches it (ADR-0134, proposed).
GroupWhat the agent must learnData from people
PerceptionFind what it can act on. Keep contact under control. Stay upright.Egocentric video, gaze, fingertip contact force, foot pressure, contact sound, head IMU
ActionPut the right force on an object at the right time, with hands and feet. Carry and balance its weight.Hand pose, per-finger contact force, foot contact and centre of pressure, body IMU
Task structurePut steps in order. Find a mistake and repair it. Predict what an object does.Step labels, task narration, error and recovery labels, outcome labels with the force trace
Force regulationUse as much force as the task needs. Keep a margin from hazards.Contact force, near-miss events, hand-to-hazard clearance
Working with peopleHand objects over. Keep a safe distance. Take turns.Handover force and release timing, gaze to the handover point, distance and heading
02The body map

From the hand
to the whole person.

The kit design starts at the bare hand and grows toward the whole person at physical work. Each channel adds a body site to one clock, and the programme adds no new device family.

The body sites of the kit designA standing figure with numbered markers: 1 at the fingertips and the wrists, 2 at the feet and the trunk, 3 at the eyes, 4 at the wrists and 5 at the mouth. The list beside the figure gives each rank as text.
Concept illustration. Each marker is a planned body site in the kit design, by rank. No channel is built or recorded.
  1. Bare-hand fingertip forcein design
  2. Foot pressure and body IMUresearch
  3. Gaze on the force clockin design
  4. Hand contact soundresearch
  5. Spoken task narrationresearch
03The ranked stack

Five channels.
In this order.

The rank comes from published research, not from purchases. A Tier 2 channel is offered to a lab only after its bet passes and a labelled channel with a validity spec exists. Tier 1 waits for the gates of its product rows.

  1. Bare-hand fingertip force

    in design

    Per-finger contact force from the sEMG of an sEMG/IMU band, calibrated against a fingertip-force rig. No sensor covers a finger pad during a demonstration. A fingernail force module can join a later revision of Ember if its own bet passes.

    Video shows where a hand went. It does not show how hard the hand pressed.

    Bets
    • contact-force-wrist-semg: funded, kill date
    • nail-force-module: proposed, kill date
    If the bet fails
    The fallbacks of ADR-0132 (proposed) apply, and no Tier 2 purchase starts until the stack is ranked again.
  2. Foot pressure and body IMU

    research

    Insoles are to give foot-contact events, the force of each step and aggregate centre-of-pressure features. Body IMUs give body motion.

    Humanoid methods use foot contact, and today they estimate it from video.

    Bets
    • insole-contact-module: proposed, kill date
    If the bet fails
    The feet stay a video estimate, and no insole build starts.
  3. Gaze on the force clock

    in design

    The gaze target from the Lens glasses, on the same clock as the contact force.

    Gaze alone is common in open datasets. Gaze on one clock with measured contact force is the intended difference.

    Bets
    • natural-gaze-target: funded, kill date
    • measured-vs-predicted-gaze: proposed, kill date
    If the bet fails
    Gaze stays a recorded column, offered on request and not sold. If measured gaze adds nothing, the gaze line falls back to the head camera, and the card says so.
  4. Hand contact sound

    research

    Features of contact sound from a wrist or nail piezo, with the speech bands removed on the device.

    A tap, a scrape and a slip make sounds that mark contact events.

    Bets
    • contact-sound-features: proposed, kill date
    If the bet fails
    No audio modality is registered.
  5. Spoken task narration

    research

    A silent take, then the person's commentary over the replay. The device makes the transcript, and the transcript labels the steps of the task.

    Narration gives the order of the steps, and the reason for a correction, as labels.

    Bets
    • task-narration-labels: proposed, kill date
    If the bet fails
    Narration stays teaching material for a skill capsule, and it is not a label for robot data.
04Labels on the same kit

No new sensor.
New labels.

Three label sets use the streams of the stack. A label names a physical event, for example a release onset. A label never names a mental state.

  1. Error recovery.

    Mistake, correction and recovery segments, with the force and gaze trace at the slip.

  2. Near-miss margins.

    Hand-to-hazard clearance, approach speed and near-miss events. These labels show where people take care. They are never a safety guarantee, and they never replace a safety standard.

  3. Two-person handover.

    The grip force of each person, the release onset and the gaze to the handover point. Each person in the episode gives consent.

05Deferred, and why

Not every sense.
Not every signal.

  • Taste is deferred. The kit has no wearable path for taste. A task note, for example "the person tastes the sauce", records it.
  • Smell is air, not perception. A gas-sensor array measures the air of a place. It is a research column under its own bet, never a channel for sale.
  • Physiology stays research. Heart-rate, skin and temperature features, thermal contact and EEG stay research columns. They never feed a product.
  • No mental-state stream. A label names a physical event, for example a release onset. The kit offers no stream that labels a mental state, and the head camera faces out, so the facial expression of the demonstrator is not recorded.
06Data limits

Rules before
the first session.

These rules apply to each channel before its first recording. Each person gives consent for each episode, and can withdraw it.

07The order of release, a later phase (ADR-0136)

Three tiers.
Each behind a gate.

Tier 1 is the band, the head camera, the Lens glasses on request, the fingertip-force rig for calibration and the LED flash fixture. Tier 2 adds insoles, body IMUs, a contact piezo and narration, each after its bet passes. Tier 3 stays research and is never offered as a product. Nothing is offered for sale today.