Frequently Asked Questions

An AI sleep companion is a bedside device that supports rest and wakefulness without wearables, cameras, or cloud uploads. It senses your presence and body contactlessly, keeps all data on-device, and learns your routines over time. It is distinct from a sleep pet or plushie (which offers comfort but no sensing) and from a wearable tracker (which requires nightly charging and wearing).

NETVERSE's first instantiation is Raychel — a spatial AI sleep and wake companion that blends embodied intelligence with elegant design. Raychel NODBOT is a camera-free alarm clock that monitors sleep through the air using mmWave radar and LiDAR, then wakes you with a gentle simulated sunrise. Raychel TIANMU gives your bedroom its own window to the sky with a circular 720p display.

Raychel is a spatial AI companion, not just an alarm. It decodes sleep stages in real time, wakes you at the lightest phase of your cycle with a phase-locked simulated sunrise (6500K dawn light), and builds a personalized long-term model of your habits, rhythms, and emotions.

Raychel uses non-contact millimeter-wave radar and Time-of-Flight LiDAR to detect breathing rate, body movement, and presence through the air. No rings, watches, or cameras are required — you simply place it beside your bed and sleep naturally.

No. Raychel is deliberately camera-free. Privacy is not a feature — it is a structural property of the spatial language. All sensing is non-visual, and all data stays on the device.

Yes. All sleep patterns, conversations, and memories are processed and stored locally on the device using an edge-native AI architecture (RWKV). No personal data is ever sent to the cloud, and nothing is sold or shared with third parties.

It is our founding philosophy, drawn from John Berger. Seeing — spatial, embodied perception — is the primary language of human-machine communication. We build machines that perceive presence, context, and rhythm before any word is spoken.

The Spatial Internet is the next evolutionary leap of the internet — a seamless, intelligent layer bridging the physical and digital worlds. The spatial language is its native interface: expressed through gesture, gaze, proximity, presence, rhythm, and context, so machines and humans can converse without clumsy translation layers.

Raychel NODBOT launched on Kickstarter and is available through the official NETVERSE shop at shop.netverse.cc.

Raychel is a calm, intelligent bedside companion designed to improve your sleep and support your daily rhythm. It consists of two products:

• Raychel NODBOT – An AI‑powered alarm clock with non‑contact sleep monitoring, gentle wake‑up light, and environmental sensing.
• Raychel TIANMU – An optional projector that connects to NODBOT and displays relaxing, AI‑generated visuals on your bedroom ceiling.

Raychel is not a general‑purpose smart speaker. It focuses on rest, privacy, and unobtrusive assistance.

• No Camera – Raychel uses only radar, LiDAR, and microphones. No image or video capture.
• Non‑Contact Sleep Tracking – No wearables. Sensors work through bedding.
• Privacy‑First Architecture – All sleep data stays on the device by default. Cloud sync is opt‑in.
• Respectful Interaction – Quiet at night, gentle in the morning, and minimal during the day.

No. NODBOT works perfectly on its own. TIANMU is an optional enhancement for users who enjoy immersive, ceiling‑projected visuals during wind‑down or sleep.

Attribute · Detail

Dimensions · 13.5 × 11.5 × 8.0 cm

Weight · 260 g

Display · 4‑inch E‑ink, 117,000 effective pixels, 4‑level grayscale

Network · Wi‑Fi 802.11b/g/n (2.4 GHz), Bluetooth 5.2

Power · 5V DC via USB‑C (adapter not included); coin cell battery for clock backup

Microphones · 2 with noise reduction

Speaker · 1, 180 Hz low‑frequency‑optimised driver, sealed BASS chamber

LEDs · 12 full‑spectrum SunLike LEDs

Buttons · 1 AI Button (multi‑function); 1 hidden RESET for factory reset

Ports · 1 USB‑C (16‑pin, supports data and power delivery)

Night‑time optimisation · Night‑light colour temperature 2700–3000K; wake‑up light up to 6500K; display remains dark; 4‑level grayscale visible in dim light; audio softened by sealed chamber

Housing · PC, ABS, resin, rubber, American walnut; Matte White / Deep Bondi Blue finish

Sensors · 60 GHz mmWave radar, 24 GHz mmWave radar, TOF LiDAR, ambient light, temperature, humidity, capacitive touch

Attribute · Detail

Dimensions · 15.0 × 13.5 × 13.5 cm

Weight · 650 g

Display technology · Circular DLP, 720p @ 60 Hz

Brightness · 165 ANSI lumens

Contrast ratio · 1200:1

Connectivity · USB‑C (to NODBOT or separate power); Wi‑Fi 6 for content updates

Projection angle · 120° tilt

Cooling · Ultra‑quiet fan (inaudible at night)

Colour temperature range · 2700K – 6000K

Operating system · Linux with Unity 3D real‑time rendering engine

Power supply · External 100–250V AC adapter (not included)

Raychel NODBOT

• NODBOT main unit

Raychel TIANMU (sold separately)

• TIANMU projector unit
• USB‑C to USB‑C cable

Power adapters are not included to reduce environmental impact. Any standard USB‑A or USB‑C adapter rated 5V/1A or higher works.

1. Place NODBOT on your bedside table (see section 4 for placement guidance).

2. Connect the USB‑C cable to NODBOT and to a power adapter (5V/1A or higher).

3. Wait for the blue LED to start blinking (approximately 10–15 seconds).

4. On your phone or computer, join the Wi‑Fi network named NODBOT-1234.

5. Follow the on‑screen instructions to connect NODBOT to your 2.4 GHz Wi‑Fi network.

6. Once connected, the blue LED becomes solid for a moment and then fades. Setup is complete.

No smartphone app is required for basic functions. An optional companion app will be available in the future for detailed data visualisation.

Simply connect TIANMU to NODBOT using a USB‑C cable. NODBOT automatically detects the projector and begins sending visual content. No pairing, no additional configuration. You may also power TIANMU separately by connecting it to its own USB‑C adapter, but connection to NODBOT is required for content generation.

Yes. Core functions — alarm, sleep detection, night light, gesture snooze — operate offline. Voice recognition and cloud‑connected personalisation require Wi‑Fi.

To reduce electronic waste. Most users already own a compatible USB‑A or USB‑C adapter. We encourage reusing existing chargers.

For optimal performance:

• On a level bedside table.
• Slightly above mattress level (e.g., on a typical nightstand).
• Facing the bed directly, with an unobstructed view of the sleeping area.
• At least 20 cm (8 inches) from the mattress edge to allow radar and LiDAR to cover the full lying area.

Not recommended. The sensors work best when the device is at roughly mattress height or slightly higher, aimed toward the centre of the bed. A high shelf reduces accuracy for posture detection and fine gestures.

No. Millimetre‑wave radar and LiDAR penetrate thin fabrics such as sheets, blankets, and curtains. However, avoid placing metal objects (e.g., metal water bottles, jewellery boxes) directly in front of the device, as they may cause signal reflections.

Yes. The radar and LiDAR systems can distinguish two individuals based on position, movement patterns, and breathing frequency. After 3–5 nights of learning, separate sleep summaries are provided for each person.

Place TIANMU on the bedside table at the same height as NODBOT, or on a low shelf. Aim the lens toward the ceiling above the bed. The short‑throw design produces a clear image even in small bedrooms. The projector can be tilted up to 120°.

NODBOT is designed for minimal, intuitive interaction. Physical controls are managed through the single AI Button located on the back.

Action · Function

Single click · Confirm / continue / dismiss a prompt

Double click · Exit / cancel current action / refresh the E‑ink screen

Press and hold (2 seconds) · Enter settings menu (time, alarm, sound, display)

In addition, you can use voice commands and hand gestures.

Examples of typical voice interactions:

• “Set an alarm for 8:20 AM.”
• “Show yesterday’s sleep record.”
• “Start a 15‑minute meditation.”
• “Turn off the night light.”
• “How many days until Christmas?”
• “Enter sleep mode with a forest scene.”
• “Switch temperature to Fahrenheit.”

At night, the response is intentionally softer and shorter to avoid fully waking you. You may also disable voice recognition entirely in settings.

When the morning alarm rings, simply raise your hand gently within about 30 cm (12 inches) of NODBOT. The mmWave radar detects the gesture and activates snooze (9 minutes by default). No button pressing, no fumbling. This works even if you keep your eyes closed.

Yes. Depending on the current mode:

• Waving left or right – adjust lighting brightness.
• Holding hand still for 2 seconds – pause or resume audio playback.
• Double wave – activate “goodnight” mode (lights dim, soundscapes begin).

You can enable or disable gesture controls via Settings → Interaction → Gestures.

From the main home screen, perform a short‑then‑long press: press the AI Button briefly, release, then press and hold for 10 seconds. The device will shut down. To restart, unplug and reconnect the USB‑C power.

Press and hold the hidden RESET button (located next to the USB‑C port) for 15 seconds. All local data — sleep records, Wi‑Fi credentials, and preferences — will be erased. The device will restart in setup mode.

NODBOT features a 4‑inch E‑ink (electronic paper) display with 117,000 effective pixels and 4 grayscale levels. E‑ink reflects ambient light like paper; no backlight is required, making it highly readable in dim bedrooms without emitting disruptive light.

Refer to the on‑device layout. The home screen includes these elements:

Element · Description

Real‑time temperature & humidity · Current room conditions. Voice control can switch between Celsius and Fahrenheit.

Date & weekday · Displayed in abbreviated format.

Raychel’s affection level · A subtle indicator showing how well Raychel has learned your preferences over time. It influences minor behavioural details.

Network status · Icon visible when connected; hidden when offline.

Holiday information · Shows today’s holiday most relevant to you (e.g., “Thanksgiving”, “Diwali”).

Current time · 24‑hour format.

Alarm status · Displays the next scheduled alarm. Hidden if no alarm is active. For workday‑only alarms, a rest icon appears on recognised holidays.

New message alert · Temporary notification (e.g., from the future companion app). Not persistent.

Raychel’s live emotion · A small animated expression that changes with each screen refresh, reflecting Raychel’s “mood”.

Text box · Shows real‑time interaction prompts, motivational quotes, or occasional thoughtful messages.

Yes. Press and hold the AI Button to enter Settings → Display. From there you can select which information widgets appear, set the display to turn off completely during defined sleep hours, and adjust refresh behaviour.

By default, the E‑ink refresh pauses during your scheduled sleep window, showing a static, dim time. You may also schedule a complete display shutdown (e.g., 11 PM to 7 AM) to further reduce any residual light.

Sensor · Primary function · Specific parameters detected

60 GHz mmWave radar · Cardiorespiratory monitoring · Breathing rate (0.1–0.8 Hz), heart rate (0.8–3.0 Hz), micro‑movements (chest wall displacement), presence

24 GHz mmWave radar · Gross motion & fallback · Large body movements (turning, getting up), wide‑area presence confirmation

TOF LiDAR · Posture & spatial mapping · Body position (supine, prone, left/right lateral), centre of mass, in‑bed/out‑of‑bed state

Temperature sensor · Environmental quality · Room temperature (continuous)

Humidity sensor · Environmental quality · Relative humidity (continuous)

Dual microphones · Acoustic event detection · Snoring, breathing sounds, ambient noise, silent gaps (potential apnoea indicators)

Ambient light sensor · Display & lighting adaptation · Adjusts E‑ink and LED behaviour based on room brightness

Millimetre‑wave radar emits low‑power radio waves that reflect off your body. The reflected phase changes with chest wall displacement caused by breathing and heartbeat. NODBOT analyses these phase changes using advanced algorithms (EWT‑VMD decomposition and TME template matching) to separate breathing and heart rate signals. This works through bedding and does not require any physical contact.

Yes. By combining breathing irregularity, heart rate variability, and micro‑movement patterns, the system estimates light, deep, and REM sleep stages. The algorithm has been validated against consumer wearables. However, Raychel is not a medical device; sleep data is for wellness purposes only.

The radar and LiDAR sensors can distinguish individuals based on spatial position and distinctive movement signatures. After an initial learning period (3–5 nights), the system provides separate sleep summaries for each side of the bed. You can also manually assign bed sides in the settings.

No. The dual microphones are used only for:

• Voice command recognition (after the wake word).
• Snoring and breathing pattern analysis.

Audio is processed locally in real time. Raw audio is never stored, uploaded, or transmitted. You may physically mute the microphones via Settings → Privacy → Mic Off.

By default, no. All sleep records, memory, and preferences remain locally on NODBOT’s internal storage (16 GB). Cloud upload occurs only if you explicitly opt in — for example, to sync with a second unit or to enable advanced personalisation. Even then, only aggregated insights (e.g., sleep scores, trends) leave the device. Raw radar, LiDAR, or microphone data never leaves your home.

Yes. The 60 GHz and 24 GHz radars operate at an average output power below 0.5 mW — significantly lower than a smartphone’s cellular radio. The TOF LiDAR is Class 1 eye‑safe. All sensors comply with FCC, CE, and other international safety standards.

NODBOT incorporates 12 SunLike LEDs that reproduce a spectrum close to natural sunlight. Available light modes:

• Night light – 2700–3000K, warm amber, ideal for winding down.
• Wake‑up light – Simulated sunrise from 2700K to 6500K over a programmed duration (10–30 minutes).
• Ambient glow – Soft, diffused illumination suitable for reading or relaxing.
• Screen illumination – Indirect light that makes the E‑ink display readable without direct backlight.

After you set an alarm, NODBOT begins the sunrise cycle 20 minutes before the alarm time (duration adjustable). The light gradually brightens and shifts from a deep orange to a cool morning white. This natural progression helps suppress melatonin and ease you into wakefulness. The alarm sound follows after the light reaches full brightness.

Yes. While awake, single‑click the AI Button to cycle through: Off → Night light (dim) → Reading light (brighter, upward‑directed) → Ambient glow. The reading light is designed to illuminate the room without shining directly into your eyes.

• Speaker – Single driver with a sealed BASS chamber tuned for soft, non‑startling night audio. Frequency response optimised at 180 Hz for pleasant low frequencies at low volumes.
• Audio library – Pre‑loaded white noises (ocean, rain, forest, fireplace), nature soundscapes, and short guided meditations (5, 10, 15 minutes).
• Wake sounds – Gentle chimes, soft piano, birdsong. No abrupt or harsh alarm tones.
• Volume – Limited to bedroom‑appropriate levels (maximum 85 dB) to prevent auditory startling.

Currently, NODBOT does not support Bluetooth streaming or third‑party music services. You will be able to upload custom audio files (e.g., personal white noise tracks) via the future companion app.

TIANMU transforms your ceiling into a canvas of AI‑generated visuals:

• Wind‑down scenes – Slow nebula motions, floating leaves, gentle ocean waves.
• Sleep‑stage art – Abstract patterns that evolve as you move through light, deep, and REM sleep.
• Dream reconstruction (experimental) – After several nights, Raychel creates a unique animation based on your movement and breathing history.
• Wake‑up visuals – An expanding circle of light that brightens in sync with the sunrise alarm.

All content is generated in real time by NODBOT’s AI. No pre‑recorded loops are used.

TIANMU includes 200+ built‑in scenes and digital art, continuously updated. Examples include: ocean, forest, bonfire, coffee_shop, crickets, keyboard, old_fan, rain, stream, train_ride.

TIANMU follows NODBOT’s active mode and responds to both settings and gestures:

Feature · Description

Screen angle · Because the projection uses a circular design, you can rotate the image to any comfortable viewing angle using voice or NODBOT controls. The angle is saved after device restart.

Focus adjustment · Non‑automatic. Use the NODBOT settings menu to assist motorised focus. It follows your hand gesture distance.

Image correction · After each movement, the image auto‑levels and corrects display distortion.

Auto activation · The projection turns on or off automatically based on user presence and spatial position (e.g., turns off when no one is nearby).

Hand interaction · Hand tracking lets you control interactive effects – for example, a bullet time effect where the image responds to your hand movements in slow motion.

Gesture switching · Wave your hand in front of the sensor to trigger an automatic switch to the next projection scene.

AI generation · Describe the visual you want using voice input. AI dynamically generates that projection effect. Optional subscription may be required.

Sleep shutdown · After the system detects that you have fallen asleep, the display gradually dims and then turns off completely. No action required.

The TIANMU hardware includes three years of AGI (AI‑generated imagery) content service. After three years, basic projection (static scenes, simple loops) continues without charge. Advanced dynamic content may require a nominal subscription. Pricing for post‑3‑year service has not yet been announced.

No. Brightness is intentionally modest (165 ANSI lumens spread over a large ceiling area). You can set TIANMU to turn off automatically after 15–60 minutes or set an ultra‑dim “star field” mode that runs throughout the night with minimal illumination.

When new content or software updates need an internet connection, TIANMU automatically syncs with NODBOT’s network connection and uses it to download and install the latest software and visuals. No manual intervention is required.

No. TIANMU is a companion device and requires a connected NODBOT to generate and control content. It does not function standalone.

Absolutely not. Raychel NODBOT contains no optical camera, no infrared camera, and no image sensor of any kind. Privacy in the bedroom is a foundational design requirement.

All sleep records, personal preferences, and long‑term memory are stored locally on NODBOT in an encrypted file system. The device includes 16 GB of dedicated storage for this purpose.

Yes. A factory reset (see section 5.6) deletes all locally stored data. You may also delete individual sleep records via the settings menu.

Cloud sync is strictly opt‑in and is disabled by default. If you choose to enable it (e.g., for multi‑device sync or advanced AI features), only the following is transmitted:

• Aggregated sleep summaries (total sleep time, sleep score, stage percentages)
• Anonymous usage patterns (e.g., “user prefers forest soundscape”)

Raw sensor data — radar signals, LiDAR depth maps, audio snippets — never leaves the device. You may revoke cloud consent at any time.

No. NODBOT does not support remote administration or inbound connections from the internet. All communication is outbound and encrypted. As with any connected device, we recommend keeping your home Wi‑Fi network secure.

Yes. The device follows privacy‑by‑default and privacy‑by‑design principles. For users in the European Union and California, optional cloud sync will include full data export and deletion capabilities. NETVERSE does not sell or share personal data with third parties.

• Ensure your hand is within 30 cm (12 inches) of the device and directly in front of the translucent radar window.
• Perform a slow, deliberate raising motion (not a fast swipe).
• Verify that gesture control is enabled: Settings → Interaction → Gestures → On.
• In a shared‑bed situation, the radar may be tracking your partner’s movement. Adjust the placement slightly toward your side.

E‑ink displays may retain faint traces of previous images. To refresh the screen completely:

• Double‑click the AI Button quickly, or
• Press and hold the button for 2 seconds, then select “Refresh Screen”.

Full refresh clears ghosting and consumes negligible power.

Adjust the volume via Settings → Alarm → Volume. If you prefer a fixed volume without gradual increase, disable the “Gradual fade” option in the same menu.

• Speak naturally, at a normal volume, within 2 metres (6 feet) of NODBOT.
• Ensure the wake word (“Hey Raychel” or “Raychel”) is enabled in Settings → Audio → Wake Word.
• Check that the microphones are not muted: Settings → Audio → Mic → On.
• In a noisy environment (e.g., loud fan or TV), try using gesture or button controls instead.

NODBOT requires 3–5 nights to learn individual movement and breathing signatures. After one week, separation accuracy improves significantly. You can also manually assign bed sides (left/right) in Settings → Sleep → Bed Assignment.

• Confirm the USB‑C cable is securely connected between TIANMU and NODBOT (or to its own power source).
• Ensure NODBOT is powered on and running the latest firmware.
• Check that the projector lens is not obstructed and that the tilt angle is set correctly.
• Disconnect and reconnect the USB‑C cable to restart the projector.

Try a force restart: disconnect the USB‑C cable, wait 10 seconds, then reconnect. If the device remains unresponsive, perform a factory reset using the hidden RESET button (see section 5.6).

• Unplug the device before cleaning.
• Use a soft, slightly damp (not wet) lint‑free microfiber cloth.
• For the E‑ink screen, a dry glasses cleaning cloth is ideal.
• Do not use alcohol, solvents, abrasive cleaners, or excessive moisture. These can damage the walnut wood finish and translucent plastic.

NODBOT is splash‑resistant. The enclosure is sealed against accidental water splashes or small drink spills. However, do not immerse the device in any liquid. If liquid enters the USB‑C port, let it dry completely before reconnecting power.

Walnut is a stable hardwood. The wood is sealed with a low‑VOC matte finish designed for typical bedroom humidity (30–70% RH). Avoid placing NODBOT in direct sunlight for extended periods, as UV exposure may darken the wood unevenly. Extreme and rapid humidity changes (e.g., next to a humidifier outlet) should be avoided.

• E‑ink display – 5+ years under normal use.
• SunLike LEDs – 50,000 hours (approximately 10 years of nightly use).
• Radar and LiDAR sensors – Solid‑state, no moving parts; lifespan exceeds 10 years.
• USB‑C port – Rated for 10,000 insertions.
• Internal storage – If the storage memory fails, NODBOT will continue to function as an alarm clock and basic sensor device, but will lose sleep history and personal memory.

Raychel NODBOT and TIANMU come with a 1‑year limited warranty covering manufacturing defects in materials and workmanship. The walnut housing and E‑ink display are covered under normal usage conditions. Damage caused by liquids, drops, or unauthorised disassembly is not covered.

• Live Chat – Available at www.netverse.cc, Monday to Friday, 9 AM – 9 PM ET.
• Email – Send your inquiry to [email protected] (response within 24 hours).
• Help Center – Video tutorials, written guides, and downloadable manuals at help.netverse.cc.

Please refer to the Return & Refund policy on our website for your region. Policies vary by country.

Sleep is not a uniform state. The American Academy of Sleep Medicine (AASM) classifies sleep into two major types: Non-Rapid Eye Movement (NREM) sleep and Rapid Eye Movement (REM) sleep. NREM is further divided into three stages: N1 (light transitional sleep), N2 (stable light sleep comprising about 45–55% of total sleep time), and N3 (deep slow-wave sleep, the most restorative stage). REM sleep is when most vivid dreaming occurs, characterized by rapid eye movements, muscle atonia, and highly active brain patterns resembling wakefulness.

A healthy adult cycles through these stages approximately 4–6 times per night, with each full cycle lasting roughly 90–100 minutes. Early cycles are dominated by deep N3 sleep; later cycles contain longer REM episodes. The proportion and continuity of each stage directly influence how restored you feel upon waking.

Why this matters for monitoring: Each stage presents a distinct physiological signature. During N3, breathing becomes slow and highly regular, heart rate drops, and body motion is minimal. During REM, breathing and heart rate become irregular, and the body enters temporary paralysis (atonia). The NETVERSE algorithm leverages these known physiological patterns — extracting respiratory regularity, heart rate variability trends, body movement intensity, and posture stability to identify sleep stages without needing EEG electrodes. Research demonstrates that using respiratory, heart rate, and body movement signals alone can achieve approximately 72% accuracy for Wake/NREM/REM classification (Cohen's kappa = 0.67); multimodal fusion in NETVERSE pushes performance significantly higher.

Recommended reading: Iber C, Ancoli-Israel S, Chesson AL, Quan SF. The AASM Manual for the Scoring of Sleep and Associated Events. American Academy of Sleep Medicine, 2007.

Polysomnography (PSG) is the reference standard for diagnosing sleep disorders. Conducted in a sleep laboratory, PSG simultaneously records multiple physiological signals: brain activity (electroencephalogram, EEG), eye movements (electrooculogram, EOG), muscle tone (electromyogram, EMG), heart rhythm (electrocardiogram, ECG), breathing airflow, respiratory effort, blood oxygen saturation (SpO2), and body position. A trained sleep technician manually scores the recording in 30-second epochs according to AASM criteria.

PSG provides the most complete picture of sleep architecture and pathology, but it has practical limitations: it requires an overnight hospital stay, involves numerous wired sensors attached to the body, and depends on expert manual scoring that is both time-consuming and subject to inter-rater variability. These constraints make PSG impractical for long-term, longitudinal monitoring or large-scale population studies.

NETVERSE's approach bridges this gap between clinical precision and everyday accessibility. By combining 60 GHz FMCW mmWave radar (for sub-millimeter chest displacement sensing), 24 GHz radar (for macro-motion detection), TOF depth sensing (for posture classification), environmental sensors, and dual microphones, NETVERSE captures the multi-signal approach of PSG — but entirely without body contact. While NETVERSE does not claim diagnostic equivalence to PSG, it aims to provide an evidence-based proxy through multimodal sensor fusion, validated against PSG-scored data from 20–30 participants in both healthy and sleep-disordered populations.

The Apnea-Hypopnea Index (AHI) is the standard clinical metric for assessing the severity of sleep apnea. It represents the average number of apneas (complete breathing pauses) and hypopneas (partial breathing reductions) per hour of sleep. According to clinical guidelines, mild sleep apnea is defined as AHI 5–14 events/h, moderate as AHI 15–29 events/h, and severe as AHI ≥ 30 events/h.

Sleep apnea affects nearly one billion people worldwide, with prevalence rising due to obesity and sedentary lifestyles. Moderate to severe sleep apnea significantly elevates the risk of coronary artery disease, congestive heart failure, stroke, and cardiac arrhythmias. AHI measured by PSG is the primary diagnostic tool, but it has known limitations: it treats all respiratory events as equal regardless of duration or associated oxygen desaturation depth.

How NETVERSE contributes: The algorithm performs AHI estimation through a multimodal respiratory event detection pipeline. The 60 GHz radar continuously tracks chest wall displacement to detect periods where breathing amplitude drops below baseline for more than 10 seconds. Simultaneously, the dual-microphone array captures corresponding silent intervals (apnea) or gasping/喘息 sounds. When both modalities independently signal a respiratory event, detection confidence increases significantly. NETVERSE's deep-learning-based detector targets an accuracy of 93.60% for respiratory event detection, with Bland-Altman analysis estimating AHI agreement within ±10 events/h of PSG-measured values. This makes NETVERSE suitable for screening and long-term monitoring, though formal diagnosis should always be confirmed through clinical PSG.

Traditional sleep monitoring methods fall into two categories: contact-based (PSG, wearable wristbands, chest straps, smart rings) and non-contact (radar, camera-based, under-mattress sensors). Contactless consumer sleep-tracking devices (CCSTDs) offer several advantages: they are unobtrusive, require no user setup at bedtime, cause no skin irritation or discomfort, and are well-suited for long-term, longitudinal monitoring in natural home environments.

Research has demonstrated that contactless technologies can reliably capture heart rate, breathing rate, and sleep-wake patterns across the sleep period in community-dwelling populations, including older adults. Compared to actigraphy (wrist-worn movement sensors), first-generation CCSTDs show comparable or better epoch-by-epoch agreement for total sleep time and sleep efficiency. However, most consumer-grade contactless devices currently provide only simplified sleep-wake classification without detailed sleep staging.

NETVERSE advances this field through a multimodal sensor array purpose-built for contactless sleep analysis. Rather than relying on a single modality, NETVERSE fuses five complementary sensor streams: 60 GHz and 24 GHz mmWave radars for cardiorespiratory vital signs and motion; a TOF (Time-of-Flight) depth sensor for posture classification; temperature-humidity sensors for environmental context; and dual microphones for acoustic event detection including snoring and breathing sounds. This sensor diversity provides resilience — when one modality experiences signal degradation (e.g., radar signal dropout during large body movements), the system dynamically reweights other sensors to maintain monitoring continuity. All sensors operate without capturing identifiable images or storing raw audio, preserving user privacy by design.

Frequency Modulated Continuous Wave (FMCW) radar at millimeter-wave frequencies (typically 60–77 GHz) has emerged as a powerful technology for non-contact vital sign monitoring. The principle is elegant: the radar emits a frequency-swept electromagnetic chirp that reflects off the human chest. Chest wall displacement caused by breathing (4–12 mm amplitude) and cardiac activity (0.2–0.5 mm amplitude) modulates the phase of the reflected signal. By extracting and analyzing this phase information, respiration rate and heart rate can be derived with high precision.

The 60 GHz frequency band offers a critical advantage: its wavelength is approximately 5 mm, making it exquisitely sensitive to sub-millimeter chest movements. Compared to 24 GHz radar, 60 GHz provides superior resolution for separating the subtle cardiac signal from the larger respiratory signal. Advanced signal processing techniques such as Variational Mode Decomposition (VMD) and Empirical Wavelet Transform (EWT) further separate breathing (0.1–0.8 Hz) and heartbeat (0.8–3.0 Hz) components from the composite chest displacement waveform.

NETVERSE's implementation takes a dual-frequency approach: a 60 GHz FMCW radar serves as the primary cardiorespiratory sensor, while a 24 GHz radar provides complementary wide-area motion detection and backup vital sign coverage. The 60 GHz radar employs phase-based displacement measurement where chest motion Δx is derived from phase changes ΔΦ through the relationship Δx = (λ/4π)·ΔΦ, with λ being the 5 mm wavelength. This yields sensitivity far beyond what consumer wearables achieve. Critically, mmWave radar penetrates blankets and clothing, operates unaffected by ambient light or darkness, and contains no camera — ensuring complete visual privacy.

Heart Rate Variability (HRV) — the natural variation in time between consecutive heartbeats — is a powerful, non-invasive window into the autonomic nervous system (ANS). During NREM sleep, parasympathetic (vagal) activity dominates: heart rate slows, HRV increases, and breathing becomes deeply regular. During REM sleep, sympathetic activity surges: heart rate becomes more variable and irregular, reflecting the brain's heightened metabolic state. During wakefulness, the sympatho-vagal balance shifts toward sympathetic dominance, with heart rate rising and HRV decreasing relative to deep sleep.

HRV analysis is performed in both the time domain (e.g., SDNN — standard deviation of RR intervals; RMSSD — root mean square of successive differences, reflecting parasympathetic activity) and the frequency domain (e.g., LF/HF ratio — the balance between sympathetic low-frequency and parasympathetic high-frequency power). These metrics have been validated for assessing sleep stage transitions, detecting sleep-disordered breathing, and evaluating insomnia severity. A higher LF/HF ratio generally indicates sympathetic activation, while a lower ratio indicates parasympathetic relaxation.

NETVERSE extracts HRV parameters directly from the radar-derived cardiac signal, computing inter-beat intervals from the phase variations in the reflected 60 GHz waveform. Key HRV features — mean RR interval, SDNN, RMSSD, pNN50, and LF/HF power ratio — are calculated for each 30-second epoch and fed into the sleep staging model. This approach mirrors the HRV analysis performed in clinical PSG, but achieves it entirely without ECG electrodes. The algorithm tracks how these HRV metrics evolve across the night, using characteristic shifts (e.g., increasing SDNN and RMSSD during NREM transitions, rapid fluctuations during REM) as key discriminators for sleep stage classification.

Body position during sleep is not merely a matter of comfort — it has measurable health consequences. Supine (back) sleeping is strongly associated with worsened obstructive sleep apnea because gravity promotes airway collapse and tongue retraction. Side sleeping, particularly on the right side, is generally associated with better sleep quality. Left-side sleeping with the head elevated is recommended for reducing gastroesophageal reflux disease (GERD) symptoms, as this position keeps the stomach below the esophagus. Neck pain is linked with asymmetrical side postures, while lower back pain correlates with supine positioning.

Clinically, sleep posture changes can serve as indirect indicators of sleep disruption — frequent position shifts often signal respiratory events, discomfort, or sleep fragmentation. Monitoring posture patterns over time provides valuable context for understanding sleep quality.

NETVERSE incorporates posture sensing through a TOF (Time-of-Flight) depth sensor with a 65° diagonal field of view. This sensor captures absolute distance information across multiple detection zones, enabling classification of four primary sleep postures: supine, prone, left lateral, and right lateral — plus absence detection for bed-exit events. The TOF sensor preserves privacy by recording only depth values, not images. Additionally, NETVERSE employs an innovative humidity fluctuation analysis method: the environmental humidity sensor detects rapid moisture changes associated with body position shifts, using the Pruned Exact Linear Time (PELT) algorithm for change-point detection. This provides a secondary, orthogonal posture-change signal that complements the depth-based posture classifier.

The bedroom environment directly influences sleep onset, depth, and continuity. A comprehensive review in Building and Environment identified evidence-based thresholds: noise should be kept below 35 dB; ambient temperature should range between 17–28°C (62–82°F) at 40–60% relative humidity; complete darkness is optimal, with blue-spectrum light particularly avoided before and during sleep; and good air quality with adequate ventilation supports uninterrupted breathing.

The physiological mechanism is clear: core body temperature must drop by approximately 1–2°C to initiate and maintain sleep. An overly warm bedroom impairs this thermoregulatory process, leading to fragmented sleep. High humidity compounds the problem by reducing the body's ability to dissipate heat through sweat evaporation. Conversely, excessively dry air can irritate the respiratory tract and cause discomfort. These environmental factors create a feedback loop — poor conditions degrade sleep quality, and degraded sleep reduces the body's ability to thermoregulate effectively.

NETVERSE uniquely integrates environmental monitoring into the sleep analysis pipeline. Temperature and humidity sensors record conditions at one-minute intervals throughout the night. An attention-based gated recurrent unit (AGRU) neural network predicts sleep thermal comfort from these environmental time series, generating a comfort score on each 30-second epoch. This score is used as a contextual modifier in the sleep staging algorithm and appears in the user's sleep quality report. The system provides actionable, data-driven recommendations — for example, identifying that bedroom temperature consistently exceeds 24°C correlates with increased wake epochs, suggesting that cooling adjustments could improve sleep continuity. NETVERSE thus connects environmental conditions to their measurable impact on sleep architecture.

A single sensor can only capture one aspect of a complex physiological state. Multimodal sensor fusion — the process of combining complementary signals from different sensor types — has consistently demonstrated superiority over unimodal approaches in sleep analysis. Recent research shows that multi-level sensor fusion can improve sleep stage classification accuracy from approximately 72% (single-modality) to over 90%, with transition detection sensitivity reaching 98% under challenging conditions. Fusion also improves robustness: when one modality experiences signal degradation, others compensate.

Two dominant fusion strategies exist: feature-level fusion (combining features from all sensors into a unified vector before classification) and decision-level fusion (classifying each modality independently, then combining predictions). The XSleepFusion framework demonstrated that a dual-stage information bottleneck approach — first compressing within-modality features to remove redundancy, then fusing across modalities with attention mechanisms — achieves superior generalization across diverse sleep datasets and signal quality conditions.

NETVERSE implements a three-tier fusion architecture. Tier 1 (Feature-Level): Approximately 47 features spanning respiratory, cardiac, motion, posture, environmental, and acoustic domains are extracted from all five sensor streams for each 30-second epoch. A multi-scale convolutional network with evolutionary attention (EAT-Net) processes these features at multiple temporal resolutions. Tier 2 (Context-Aware Fusion): Dynamic reweighting automatically adjusts sensor contributions — during body movement events when radar signal quality temporarily drops, TOF posture sensing and microphone input gain weight; during quiet sleep when radar SNR is high, cardiorespiratory features dominate. Tier 3 (Decision-Level): A Hidden Markov Model (HMM) performs temporal smoothing on the fused predictions, using Viterbi decoding to ensure biologically plausible state transitions. This three-level design provides both precision and resilience.

Sleep Latency (SL) — the time it takes to transition from "lights out" to the first epoch of sleep — is a fundamental measure of sleep drive and pre-sleep arousal. In healthy adults, normal sleep latency ranges from 10 to 20 minutes. Prolonged sleep latency (>30 minutes) is characteristic of insomnia and often reflects heightened cognitive or physiological arousal ("trying too hard to sleep"). Abnormally short sleep latency (<5 minutes) may indicate excessive daytime sleepiness, potentially signaling sleep apnea, narcolepsy, or chronic sleep deprivation.

Sleep Efficiency (SE) — the ratio of total sleep time to total time spent in bed, expressed as a percentage — reflects sleep continuity. A SE above 85% is generally considered normal; values below 80% suggest fragmented sleep with excessive wake-after-sleep-onset (WASO) time. Both metrics are clinically used in the Pittsburgh Sleep Quality Index (PSQI) and are sensitive to many sleep disorders as well as environmental disruptors like noise, temperature, and light.

How NETVERSE calculates these metrics: Sleep onset is determined by the first sustained epoch where the fused multimodal classifier transitions from Wake to N1 or N2, confirmed by characteristic changes in respiratory regularity, heart rate deceleration, and cessation of voluntary movement. Total sleep time is accumulated across all NREM and REM epochs; wake time is summed from all epochs classified as Wake after sleep onset. NETVERSE reports sleep latency, sleep efficiency, WASO, number of awakenings, and a hypnogram visualization — all derived non-invasively, without requiring the user to wear any device or manually log sleep times.

Stage N3, also called slow-wave sleep (SWS) or deep sleep, is the most restorative sleep stage. During N3, the brain generates high-amplitude, low-frequency delta waves (0.5–4 Hz). Growth hormone is secreted, promoting tissue repair, muscle growth, and bone density maintenance. The glymphatic system — the brain's waste clearance mechanism — becomes highly active during deep sleep, flushing out metabolic byproducts including beta-amyloid, a protein associated with Alzheimer's disease.

Physiologically, N3 is characterized by the most stable autonomic state: heart rate reaches its nightly minimum, blood pressure drops (nocturnal dipping, which is cardioprotective), breathing becomes slow and highly regular, and skeletal muscle tone is maximally relaxed but not paralyzed. Body temperature regulation remains intact, unlike in REM. Arousal threshold is highest during N3 — it is hardest to wake someone from deep sleep — and if awakened, individuals typically experience significant sleep inertia (grogginess). Deep sleep predominates in the first half of the night and decreases with age; by age 60, many adults experience substantially reduced N3.

NETVERSE identifies N3 through a characteristic physiological profile: respiratory rate at its lowest and most regular (coefficient of variation <10%), heart rate at its nightly minimum with high HRV (elevated RMSSD and pNN50), body movement approaching zero, and stable posture maintained over consecutive epochs. The bidirectional LSTM within the classification model learns to recognize this multi-signal N3 signature, achieving high specificity for deep sleep detection. The sleep report quantifies N3 duration and percentage, providing users with a meaningful window into their restorative sleep.

REM sleep is a paradoxical state: the brain is metabolically as active as during wakefulness, yet the body (below the eyes and diaphragm) is nearly paralyzed through motoneuron inhibition. This state, discovered by Aserinsky and Kleitman in 1953, is when the most vivid, narrative dreams occur. REM accounts for approximately 20–25% of total sleep in adults, concentrated in the later sleep cycles of the night.

REM serves critical cognitive functions. It facilitates memory consolidation — particularly procedural memory (skills, habits) and emotional memory processing. During REM, the brain replays and reorganizes learned experiences, strengthening adaptive memories while weakening irrelevant ones. REM deprivation impairs learning, emotional regulation, and creative problem-solving. Physiologically, REM is characterized by irregular heart rate and respiration (with transient cardiac and respiratory instability), complete loss of thermoregulation (body temperature drifts toward ambient), and phasic bursts of rapid eye movements linked to ponto-geniculo-occipital (PGO) wave activity from the brainstem.

NETVERSE detects REM through a distinctive combination of features: heart rate irregularity with transient tachycardia interspersed with bradycardia, respiratory rate variability (most irregular of all stages), near-absence of voluntary body movement but presence of small phasic muscle twitches detected by radar, and — when combined with TOF data — sustained posture with minimal position changes. The volatility-driven hidden state modeling (VDSM) framework within NETVERSE is particularly effective for REM detection because GARCH (Generalized Autoregressive Conditional Heteroskedasticity) volatility analysis captures the unique signal instability characteristic of REM transitions.

Snoring is not just a social nuisance — it is a vibration of upper airway soft tissues indicating partial airway narrowing during sleep. While occasional, low-intensity snoring may be benign, habitual loud snoring is a cardinal symptom of obstructive sleep apnea (OSA). Acoustic analysis can distinguish simple snoring from pathological respiratory events: frequency-domain features like Mel-Frequency Cepstral Coefficients (MFCC), spectral centroid, spectral roll-off, and short-time energy can differentiate snoring from non-snoring acoustic events with high accuracy.

Beyond snoring, the duration of silent intervals between breaths carries important clinical information. Prolonged silence (≥10 seconds) followed by a gasping sound strongly suggests an apnea event. The pattern and frequency of these events, combined with the acoustic characteristics of recovery breaths, correlate with the apnea-hypopnea index measured by PSG. Deep learning models such as SoundSleepNet have demonstrated that audio-based sleep breathing event detection can approach the performance of traditional respiratory monitoring.

NETVERSE's dual-microphone array captures these acoustic signatures with spatial awareness. Using Time Difference of Arrival (TDOA) between the two microphones, the system localizes sound sources — distinguishing the user's snoring and breathing from ambient noise or a partner's sounds. Machine-learning classifiers process extracted acoustic features to identify snoring episodes, breathing pauses, and abnormal respiratory sounds. Crucially, NETVERSE performs cross-modal validation: when the microphone detects a potential apnea (silence followed by gasping), the system simultaneously checks the radar-derived chest displacement signal for a corresponding period of absent respiratory motion. Only when both modalities agree does the event receive high confidence. All acoustic analysis is performed at the edge; raw audio is never stored or transmitted, only feature vectors used for classification are retained.

The World Health Organization recognizes sleep as fundamental to health. Quality sleep supports immune function — during deep sleep, the body produces cytokines that help combat infection and inflammation. Chronic sleep deficiency impairs the immune response to vaccines and increases susceptibility to respiratory infections. Sleep also regulates metabolism: inadequate sleep disrupts ghrelin and leptin balance, increasing hunger and cravings while reducing insulin sensitivity, contributing to obesity and type 2 diabetes risk.

Cognitive function is deeply sleep-dependent. The brain consolidates memories during sleep through hippocampal-neocortical dialogue, transferring fragile short-term memories into stable long-term storage. Sleep deprivation impairs attention, working memory, decision-making, and creative problem-solving. Epidemiological data shows that individuals who habitually sleep less than 6 hours or more than 9 hours face elevated all-cause mortality risk — the relationship is U-shaped, with 7–8 hours associated with optimal longevity outcomes. There is a 2–4 times greater risk of developing major depression among individuals with chronic insomnia.

NETVERSE supports health awareness by translating complex physiological data into actionable sleep insights. Rather than simply reporting a single "sleep score," NETVERSE provides a multi-dimensional breakdown: sleep architecture (stage proportions and hypnogram), respiratory quality (estimated AHI, snoring time percentage, breathing irregularity index), movement patterns (posture shifts, motor activity distribution), environmental quality (temperature, humidity, noise level trends), and a composite sleep quality score weighted across these dimensions. This comprehensive view helps users understand not just how much they slept, but how well — and what specific factors may be affecting their sleep.

Sleep hygiene refers to the set of behavioral and environmental practices that promote consistent, restorative sleep. Key recommendations include:

• Maintain consistent sleep-wake timing — going to bed and waking at the same time every day, including weekends, stabilizes the circadian rhythm.
• Optimize the bedroom environment — keep temperature at 17–22°C (62–72°F), humidity at 40–60%, light at complete darkness, and noise below 35 dB.
• Limit evening light exposure — blue-enriched light from screens suppresses melatonin secretion and delays sleep onset; avoid screens 1–2 hours before bed or use blue-light filters.
• Avoid caffeine, nicotine, and alcohol close to bedtime — caffeine has a half-life of 3–7 hours; alcohol may induce sleep initially but fragments sleep in the second half of the night.
• Exercise regularly, but not within 2–3 hours of bedtime — regular physical activity increases deep sleep duration.
• Use the bed only for sleep — associating bed exclusively with sleep strengthens the conditioned sleep response.

NETVERSE provides objective feedback on how well these practices are working. Users can observe correlations between their behaviors and sleep metrics over time — for example, noticing that nights following afternoon exercise show increased N3 (deep sleep) percentage, or that bedroom temperature above 24°C correlates with more awakenings. The system does not make prescriptive medical claims but enables data-informed self-discovery about what conditions optimize an individual's sleep.

Artificial intelligence is transforming sleep medicine. The manual scoring of a single PSG recording takes a trained technologist 2–4 hours and is subject to approximately 15–20% inter-rater variability. Deep learning models can now perform automated sleep staging in minutes with accuracy approaching expert consensus. Hybrid architectures combining convolutional neural networks (for multi-scale feature extraction) with bidirectional LSTMs (for temporal context modeling) and attention mechanisms (for learning long-range dependencies) represent the current state of the art.

The XSleepFusion framework, validated on over 15,000 recordings across eight clinical datasets, demonstrated that a dual-stage information bottleneck approach can achieve superior generalization. The Evolutionary Attention Transformer (EAT-Net) backbone processes temporal features at multiple scales while producing interpretable attention patterns that show which signal features the model weights most heavily for each sleep stage classification — an important step toward explainable AI in healthcare.

NETVERSE employs a carefully designed AI pipeline. The VDSM (Volatility-Driven Hidden State Modeling) framework combines GARCH volatility modeling (capturing intra-stage signal variability) with Hidden Markov Model inference (ensuring biologically plausible state transitions). A multi-scale CNN extracts features at temporal resolutions from seconds to minutes; a bidirectional LSTM models the sequential context of sleep stage progression; and multi-head self-attention layers learn which modalities and features matter most at each moment. The entire pipeline was architected for edge deployment — radar signal processing and feature extraction can run on embedded hardware, while the deep learning inference can execute on a local edge AI accelerator or in the cloud. This enables real-time sleep analysis without requiring continuous internet connectivity.

Sleep tracking technologies span a spectrum from clinical (PSG) to consumer (wearables and contactless). Each approach has distinct trade-offs:

Approach · Comfort · Sleep Stage Accuracy · Long-Term Viability · Privacy

PSG (clinical) · Low (many wires) · Gold standard · Single night only · High (clinical data)

Wearables (watch, ring) · Moderate (on-body) · ~60–75% for 4-stage · Good (daily use) · Moderate

Under-mattress · High (invisible) · ~55–70% for sleep/wake · Excellent (passive) · High

Contactless radar+ · High (no contact) · ≥80% targeted (5-stage) · Excellent (passive) · High (no images)

Wearables offer portability but require nightly charging, can cause skin irritation, and their optical heart rate sensors (photoplethysmography/PPG) are susceptible to motion artifacts and skin tone variability. Contactless approaches eliminate these burdens, making them particularly valuable for elderly users, children, and anyone who finds wearable devices intrusive during sleep.

NETVERSE's position in this landscape is that of a non-contact system delivering multi-dimensional sleep analysis — not just sleep-wake classification. The combination of mmWave radar (for clinical-grade cardiorespiratory sensing), TOF depth sensing (for posture analysis), environmental monitoring (for context), and acoustic analysis (for respiratory event detection) provides a richer data stream than any single-wearable device can offer. NETVERSE is not a replacement for diagnostic PSG but aims to provide a meaningful, scientifically-grounded window into sleep health for everyday use — bridging the gap between simplified consumer trackers and the inaccessibility of clinical sleep labs.

This knowledge base is part of the NETVERSE Multimodal Non-Contact Sleep Monitoring Algorithm documentation. For technical specifications, validation methodology, and deployment guidelines, please refer to the full NETVERSE technical documentation.