NODBOT · ALGORITHM NOTES
Two algorithms run the day. A spatial-sensing layer reads the room and decides what is happening in it. A sleep-monitoring layer reads the body through a blanket. Together they turn a stream of weak reflections into one confident action — a light that rises, a screen that changes, a morning that begins gently.
01 · SYSTEM
The product is built as four stacked layers. Each layer answers a narrower question than the one below it, so a decision at the top can always be traced back to evidence at the bottom.
02 · PART ONE
Interaction starts long before you touch anything. The question is not “was a pixel crossed?” but “is there a person, is that person in bed, and is that person asleep?” The system answers it by accumulating independent evidence instead of trusting one sensor.
Rather than a single distance reading, the depth array reports a multi-zone depth field. Each zone carries its own depth and motion, so a bed, a chair and a doorway do not collapse into one number. The ToF module covers a 65° diagonal field of view up to 200 cm.
A high-frequency stream preserves the small chest displacement that breathing produces. A low-frequency stable stream is built from a per-zone median and a high quantile, which removes transient spikes without erasing real change. A zero-copy single-slot queue keeps memory flat over a whole night.
When the room is empty, several snapshots per zone are taken and reduced to a per-zone median, then frozen. From that moment, a person is not “something detected” — a person is a deviation from a known-empty reference. This is what separates a sleeper from bedding, a chair or a bag.
Presence is the union of three independent paths: motion energy, relative ambient-sound change, and radar presence when fitted. A majority vote over a second-scale window confirms it. Leaving requires tens of seconds without a valid trigger, so a blanket falling off the bed cannot end the night.
In-bed detection is an AND, not an OR. Depth geometry, radar in-position and zone reinforcement establish occupancy; a vital-sign layer above the empty baseline must then prove that the object is alive. A latch, renewed by real movement, holds the state while a brief stillness would otherwise look like departure.
When “acoustic active” and “motion active” coincide persistently, the system recognises a television, an air conditioner or a fan and stops the acoustic dimension from voting. The design principle is blunt: better to abstain than to vote wrongly.
“Present” never silently becomes “in bed”, and “in position” never silently becomes “focused”. Each rung of the state ladder requires its own independent evidence. The baseline also switches between a cold-start mode and a high-precision mode as valid zones settle.
Behaviour follows a fixed escalation order: light, then screen, then tone or soundscape, then voice, then device linkage. If a soft light can answer the moment, the speaker stays silent. At night the device records; on a toilet trip it lights only the floor and dims again on return.
The minimum viable evidence set. Enough to know a room is occupied and to drive light and screen behaviour without a radar module.
Each hardware tier plugs into the same state ladder. Fusion is a change of evidence weights, not a separate code path: basic adds standard radar, advanced adds high-precision radar for heart and breathing signals.
03 · PART TWO
A radar does not see a body. It sees a phase that moves. The monitoring pipeline turns that phase into breathing, heartbeat, posture and movement, then fuses them into a five-stage night — thirty seconds at a time, the same epoch the clinical standard uses.
Radar supplies vital signs, ToF supplies posture and position, microphones supply breath and snore evidence, and the environment sensor supplies thermal context. No single one is trusted alone.
A range FFT finds the sleeper's range bin. The phase is unwrapped and converted to displacement. EWT-VMD separates the breathing and cardiac components far more cleanly than a single bandpass filter, which is what makes a one-second heart-rate estimate possible.
Features are computed on 30-second epochs — the same window used in sleep scoring — giving a vector of roughly 47 dimensions across respiration, cardiac, movement, posture, environment and acoustic groups.
ToF zones and their centres of mass classify supine, prone, left lateral, right lateral and absent. A multi-scale CNN, BiLSTM and multi-head attention stack turns the epoch vector into five-stage probabilities.
During movement, when radar signal-to-noise falls, ToF and audio gain weight. During quiet sleep, radar dominates. During snoring, an acoustic–radar consistency check is switched on. Feature-level, context-level and decision-level fusion all operate at once.
An HMM with five hidden states — Wake, N1, N2, N3, REM — is decoded by Viterbi across the entire night, so a single noisy epoch cannot break the sequence. Volatility modelling (GARCH) further stabilises abrupt transitions.
A breathing event is flagged when radar respiration amplitude stays below baseline for more than ten seconds and the microphone agrees with silence or a gasp. Two-path agreement raises confidence; an estimated AHI is derived, never presented as a diagnosis.
The wake window is taken from the current stage sequence plus elapsed time and ambient light. The sunrise begins inside a light-sleep window and climbs gradually; sound is layered on only if the light alone does not bring you out.
The staging targets for the product are at least 80% accuracy across five stages, at least 88% across three stages, a Cohen's kappa of at least 0.75, and an AHI Bland–Altman 95% limit of agreement within ±10 events per hour. Validation is designed around polysomnography: 20–30 participants over one to two nights, epochs scored by sleep technicians under the AASM standard, with ablation runs for radar-only, radar plus audio, radar plus audio plus ToF, and all sensors. Published research figures cited in the design — 72% three-class staging, kappa 0.67, above 90% after fusion, 93.6% for one radar model — are literature values, not measurements of this product. NODBOT is not a medical device, and its sleep data is descriptive rather than diagnostic.
There is no camera in the product. Raw depth frames and radar waveforms are processed on the device; the cloud does not receive them. What may leave the device is limited to what a language or content task requires.
Sleep monitoring, spatial sensing and proactive service are separately disableable. Sleep records, conversations and long-term preferences are stored locally and encrypted.
04 · CLOSING THE LOOP
The two algorithms share one vocabulary. The interaction layer knows whether the room is empty or occupied; the sleep layer knows whether the occupant is awake or in REM. The agent layer reads both, then chooses one channel and one intensity.
Every behaviour elsewhere on this site is produced by the pipeline above — no camera, no wearable, no guesswork.