Plone Mraz

Sensing

Per-channel envelopes, aggregate rate, and the on-body/off-body compute split

Open on GitHub Mở trên GitHub

Snapshot of this document in the GEMs repository. The repository itself is the current version. Bản chụp tài liệu này trong repo GEMs. Bản đang chạy nằm ở chính repo.

This is the primary axis: the body exists to acquire physical experience. Quantifying it forces a question that a single multiplier cannot answer.

5.1 "× human" is not one number — it is five axes

"As sensitive as human skin, times N" decomposes into five independent quantities, and they do not all move together:

Axis Human Can it exceed human
Point density ~241 mechanoreceptors/cm² at the fingertip (sourced) No — see below
Sensitivity threshold — Yes, substantially
Temporal bandwidth Pacinian corpuscles to ~500 Hz Yes — piezoelectric sensors reach tens of kHz ⇒ 20–100×
Dynamic range finite, saturates early Yes
Simultaneous modalities pressure, vibration, temperature, pain Yes — the self-healing layer of 04.2 also senses pH and humidity

Density is blocked twice, independently:

  1. Fabrication. The densest published e-skin array reaches ~347 elements/cm² (sourced) — roughly level with human skin, not above it. A 100× density means 24,100/cm², about 69× beyond anything yet built.
  2. Bandwidth. Even if it could be built: 24,100/cm² over 1.6 m² is 386 million points. At 1 kHz and 12 bits that is ~4,600 Gbps of tactile data alone — 580× the 8 Gbps link of 01. No compression ratio rescues that.

Declared tactile capability: density level with human skin in the fine regions (hands, face), thinning across the torso; 20–100× human in temporal bandwidth; above human in sensitivity threshold, dynamic range, and number of simultaneously measured quantities.

5.2 Tactile layout

Uniform whole-body density is wasteful — human skin is not uniform either.

Region Area Density Points
Fine — hands, face ~500 cm² ~300/cm² (near the fabrication ceiling) ~150,000
Ordinary — remainder ~15,500 cm² ~20/cm² ~310,000
Total 1.6 m² — ~460,000

At 1 kHz and 12 bits: 5.5 Gbps raw — two thirds of the link, from one channel. (1.8 m², ~500,000 points and 6.0 Gbps until the height came to 1.65 m on 2026-09-26.)

5.3 Channels

Channel Anchor configuration Raw rate Note
Stereo vision 2 × 4K, 30 fps, 12 bit 5.97 Gbps Optical zoom, multispectral, thermal IR retained
Thermal IR 640×480, 30 fps, 16 bit 0.15 Gbps
LiDAR / depth 300k points/s 0.04 Gbps
Microphone array 16 ch, 48 kHz, 24 bit 0.02 Gbps High-sensitivity, directional
Proprioception 41 joints × 4 channels + IMU, 1 kHz 0.006 Gbps Joint count declared at hardware/kinematics.md. Cheapest channel, and the one that must not be cut — see 5.5
SDR / multi-band RF 2 ch × 56 MHz I/Q, 16 bit 3.58 Gbps
Spatial RF sensing Wi-Fi / mmWave CSI modest Retained at full hardware capability; use is ⟦CTRL⟧
Olfaction (e-nose) — negligible 5–30 ppb achieved per compound (sourced); 5–10 s response
Taste (lab-on-chip) — negligible Slow, batch-wise
Trajectory prediction ~1.5 s horizon — Computer vision plus human trajectory prediction. TM
Voice and ultrasound — — Synthesis plus ultrasonic emission. TM
Biological response pupil, expression — Micro facial actuators and microfluidics. TM/LAB
Cardiac co-regulation vibration, warmth, rhythm — Optional module, not always-on. Instrumented per 06.6

On olfaction. Parts-per-billion thresholds are achieved. What has to be reduced is not sensitivity but sensitivity × breadth of compounds × speed simultaneously: a sensor reaching 5 ppb for one compound takes 5–10 seconds, and no array reaches ppb across hundreds of compounds in real time.

Configuration Vision Tactile SDR Total Minimum compression
A — conservative (4K30, 460k points) 5.97 5.52 3.58 15.1 Gbps 2 : 1
B — moderate (8K60, 1M points, 4-ch SDR) 47.8 12.0 7.17 67.2 Gbps 8 : 1
C — 100× density 47.8 5,206 7.17 5,261 Gbps 658 : 1 — excluded

Available link: 8 Gbps.

The on-board / off-board split of 01 is not a design preference. It falls short by 2× in the most conservative configuration and by 8× in a sensible one. On-body processing is arithmetic, not taste.

5.5 Proprioception is mandatory

Joint encoders, IMU and force/torque sensing, dense enough to distinguish "I just moved" from "something moved me".

This is the cheapest channel on the list and the one that cannot be omitted. Two separate requirements land on it:

The same hardware serves a third purpose in 06.3.

5.6 The on-body / off-body compute split

What is split here is processing, not decision. Deciding is the Main Controller's, off the body (01); the dial sets how much of the processing — compression, feature extraction, fusion — runs on the unit.

The split is a dial with two hard ends and a soft middle.

Lower end — what must be on the body:

Constraint Figure
Minimum compression 2 : 1 → 8 : 1 (5.4)
Balance loop ≥ 500 Hz
Fast reaction loop ≤ 10 ms
Context record for every output event (RSIL INV-8: anchored context on every emission) per event (08.2)

Link round-trip is ~1 ms at ideal short range, but at kilometre range with beam tracking it cannot be relied on for ≤10 ms. Balance and reflex are therefore on-body — not because the design prefers it, but because there is no alternative.

Upper end — what must not be on the body:

Constraint Content
Losing a unit loses its state Its own state and its unsynchronised record, never memory; the more processing on the unit, the more of that there is (01)
Power Edge compute draws on the same p (W/kg) that enters the convergence condition Σf < 1
Heat A sealed body in human contact dissipates worse than a rack
Memory and world model Off-body by architectural definition, not by capacity

Edge compute is not "add as much as you like, it only costs battery." It eats into the convergence condition of the whole body, and 02.3 shows γ inflating non-linearly near the ceiling.

The middle — feature extraction, compression, predictive modelling, tracking of other agents — is ⟦CTRL⟧, and should be adjustable per mission rather than a compile-time constant.

Operating point: ⟦CTRL⟧.