No register trick, I’m afraid. As far as I remember the sleep mode table in the 328P datasheet, the analog comparator is only a wake source in idle (and ADC noise reduction), not power-down. In power-down only the external interrupts, pin change, TWI address match and the watchdog wake it. Your idle current is the clocks running.

Move the comparator outside. A nanopower comparator with push-pull output into INT0 (D2) lets the 328P sit in power-down at a few µA. TLV3691 draws about 75 nA, and the reference can be a micropower 2.5 V shunt. Add hysteresis with a resistor from output to the + input, a few hundred mV at the battery side, otherwise a battery hovering at 11.8 V will wake you a hundred times a night as the load changes.

attachInterrupt(digitalPinToInterrupt(2), wake, FALLING);
set_sleep_mode(SLEEP_MODE_PWR_DOWN);
sleep_enable(); sleep_cpu();     // wakes on INT0 only
// on wake: read ADC once for the real value, send it, sleep again

Aoife: you can use the bandgap internally, but the comparator itself still can’t wake power-down, so it doesn’t help here.

On the Hermetry side I’d still send a heartbeat once a day with the voltage. Then an offline alarm on the profile tells you a sensor died, which a wake-only design never would. Put the low-voltage alarm rule on bat_v with some hysteresis too.

Did this to a pile of SG90s for a teaching rig. The mod:

1. Open the case, take out the output gear. Cut or file off the little plastic tab on its underside that hits the end stop.
2. The potentiometer under the output shaft is the position feedback. Unsolder it and put two 2.2k resistors in its place as a fixed divider (ends to the outer pads, the middle to the wiper pad). The servo now always thinks it’s at centre.
3. Reassemble. Now the pulse width sets speed and direction instead of position: about 1.5 ms stops it, shorter turns one way, longer the other. Find the exact stop value by trial, resistor tolerance shifts it.

To your second question: no, it no longer knows where it is, same as an FS90R. For a blind that matters, so do what Keisha said, add a reed switch or microswitch at fully open and fully closed, and count time or add a cheap encoder in between. Honestly, an SG90 is also weak for a real blind, a 28BYJ-48 stepper with a ULN2003 board gives you position for free and more torque.

On the board side: add blind.set with a Number parameter 0-100 to the ESP32’s profile Commands tab, Basic tier. The Slider sends the value once on release, not while you drag, which suits a motor. Have the ESP32 publish its real position as blind_pct and bind the Slider’s Current value to it, so the card shows where the blind actually is.

Power the servo or stepper from its own 5 V supply, not the ESP32’s pin.

Until something like that exists, do the zone decision on the robot. That’s also what the personnel tracking solution expects: the tracker or its gateway reports the zone name, and Hermetry’s rules work on that. I run exactly this for carts in a warehouse, ESP32 scanning fixed BLE tags.

every 2 s: scan 1.5 s, keep RSSI per known tag (EMA, alpha 0.3)
best = tag with highest smoothed RSSI
switch zone only if best beats the current zone's tag by 6 dB
             for 3 scans in a row          (stops flapping at borders)
publish on change and every 60 s:
  {"zone": "F3-east", "zone_id": 32, "zone_rssi": -71}

Send both: zone as text for tables and the alarm, zone_id as a number because text keys have no chart history and you’ll want “where was it at 14:00”.

For the kitchen corridor: give each cleaner a server attribute forbiddenZone = “F0-kitchen” and an alarm rule on the robot profile where zone equals that attribute. Per-device thresholds from attributes work for text too with ==. Same pattern the tracking solution uses, described on Personnel tracking.

Polygons: no workaround on the map that I know of. Map the polygon to tag zones instead, one zone per forbidden area.

Aoife has the core of it.

Cutting the solenoid only puts the directional valve into its rest position. Depending on the valve type that can trap air in the cylinder (closed-centre 5/3 valves do exactly that), so the cylinder can still move when someone opens a fitting, or hold a load that drops later. A stuck valve spool is the other failure: power off, valve doesn’t move. Pneumatic energy is stored energy, and for safe access you want it gone, not held. The dump valve exhausts the supply to atmosphere so the cell ends up at zero pressure. The safety relay usually also monitors the dump valve’s feedback (a pressure switch or a position contact), so a dump valve that fails to vent stops the cell from restarting. The two parts cover different failures. How much of this you need follows from the risk assessment and the required performance level under ISO 13849, which is your engineer’s job to document.

On Hermetry Safe: there’s no doc page for it, so I won’t guess what it does. Whatever it does, I would never put a cloud platform into the safety function itself. Even the personnel tracking page calls that solution an information tool, not a certified safety system. Emergency stop, safety relay and dump valve stay hardwired.

What works today and is useful: monitoring. Wire the safety relay’s auxiliary contact and the dump valve’s pressure switch into the cell PLC or a gateway and send them as telemetry, say estop.active and dump.vented. Then an alarm rule can tell maintenance when the e-stop has been active for a long time, or when the e-stop is active and dump.vented is still 0. That gives you records and notifications without touching the safety chain.