The one perception layer where "print it" is literally true: a printed magnetized-elastomer skin over a COTS magnetometer array. ReSkin, AnySkin, and eFlesh already proved it works and generalizes — so the open question is not whether, it is what whole-body coverage costs in joules, and how the ledger reshapes the architecture. The answer: fixed-rate dense sensing is ~12 W; event-driven sensing is ~0.4–1.7 W. That 7–32× gap is the entire contribution.
| Readout | µJ / sample | samples / J | mW/taxel @100 Hz | role |
|---|---|---|---|---|
| MLX90393 | 31.9 | 31 400 | 3.19 | high-res · fingertips · the ReSkin/AnySkin part |
| TLV493D | 13.7 | 72 900 | 1.37 | low-power · body presence & baseline scan |
| ASIC (fab) | 1.7 | 593 800 | 0.17 | integrated AFE+ADC+threshold · the escape |
Inside each sample, joules also buy resolution. Oversampling averages down magnetometer noise as 1/√N, and energy rises with N — so per-sample cost spans ~20×:
| oversampling | µJ / sample | noise floor | use |
|---|---|---|---|
| OSR 1 | 14.5 | ~0.60 µT | coarse — contact present / absent |
| OSR 4 | 31.9 | ~0.30 µT | slip, gross force |
| OSR 16 | 83.8 | ~0.15 µT | fine force / shear — fingertips |
| OSR 64 | 291.8 | ~0.075 µT | texture — spend only where it pays |
| Region | area | pitch | taxels | rate | chip |
|---|---|---|---|---|---|
| Hands (2×) | 0.040 m² | 5 mm | 1 600 | 200 Hz | MLX |
| Forearms (2×) | 0.060 m² | 10 mm | 600 | 100 Hz | MLX |
| Torso · thighs · shins | 0.400 m² | 25 mm | 639 | 25 Hz | TLV |
| TOTAL | 0.50 m² | — | 2 839 | — | — |
Fingertips may instead carry vision-based tactile (DIGIT-class) for micron detail; if magnetic, they sit at the dense 5 mm / high-rate end. Density and rate are the two dials — both feed the power below.
Fixed-rate dense is dominated by the hands — 1 600 taxels at 200 Hz is 10.2 W by itself. But most taxels are not in contact most of the time. Scan the whole body slowly for contact detection, and ramp sample-rate only on the taxels near an active contact — exactly the event-camera trade. Average power tracks contact activity: 7× lower manipulating, 32× lower standing. Tactile stops being a fixed 12 W tax and becomes a variable cost you only pay when touching something.
A weaker magnetized skin gives a smaller field shift per unit force, so more oversampling is needed to hit a target force resolution — which costs more joules per sample. The skin material's remanence directly sets the energy budget. That remanence is unmeasured, and it is the same coupon as MEAS-1 — just in the soft, low-loading regime. Measure Br, ΔB-per-newton, and the noise floor, and the whole OSR-and-joules column becomes real. One coupon family gates both actuator torque and tactile energy.
7–32×. Scan slow for presence, ramp fast only near contact. The single largest lever, and it is pure firmware once the patch MCU can threshold locally.
~20×. MLX (32 µJ) → TLV (14 µJ) → fab ASIC (1.7 µJ). The ASIC is the sovereignty escape: integrate the front-end so a taxel-sample stops paying I²C and conversion overhead.
~20×. Fine force costs joules. Spend OSR-64 on fingertips, OSR-1 on the torso. Do not pay texture-grade energy for presence sensing.
Pitch sets taxel count (∝ 1/pitch²); rate sets samples/s. Both linear in power. Dense-fast only where dexterity needs it; coarse-slow everywhere else.
A printed magnetized-elastomer skin, a COTS magnetometer flex array, a local patch MCU that emits events, and an event-driven central loop. Whole-body tactile then costs ~0.4 W standing and ~1.7 W manipulating — not the 12 W a naïve dense scan implies. That gap is the contribution: to our knowledge no one has published tactile skin as an energy-accounted, event-driven, CC0 system with a joules-per-taxel ledger. The working parts (AnySkin/ReSkin/eFlesh, MLX/TLV) are all open or commodity; the missing piece is the accounting and the architecture that follows from it.
Gate: measure the soft-skin Br (MEAS-1, soft regime) — it sets both the touch signal and the energy to read it. And the fab ASIC is the one place your silicon buys a 20× that no amount of firmware can.