COMO Collar

This page exclusively about COMO AI Collar (Hardware).

Modern pet owners and researchers are increasingly interested in technology to better understand and care for dogs. The trend of pet ownership among younger generations (especially Gen Z) has reshaped priorities – 76% of Gen Z prefer pets over children, spending hundreds of billions on pet care 1 . Yet, a significant gap remains in objectively understanding canine psychology and health. Unlike humans, dogs cannot verbally communicate stress, pain, or emotion; owners and veterinarians must infer these from behavior and limited vital signs. Traditional assessments of dog behavior are often subjective – even trained evaluators rely on checklists and personal judgment, which can lead to inconsistent interpretations 2 . There is a clear motivation to develop intelligent canine wearables that can continuously monitor a dog’s physical and emotional state, providing objective data and enhancing communication between pets and humans.

Wearable technology for dogs promises to fill this gap. By instrumenting a dog’s collar with sensors for heart rate, motion, temperature, sound, etc., researchers can obtain real-time insights into the animal’s well-being. This could revolutionize fields like animal behavior research, veterinary telemedicine, and everyday pet care. The COMO Collar (Canine Emotion Monitoring & Observer Collar) is designed as such a smart wearable – aiming to monitor a dog’s physiological signals and behaviors, infer emotional states through AI, and enable intelligent interaction (e.g. alerting owners or cueing the dog via vibrations). The potential impact spans pet behavior research (e.g. objectively tracking anxiety or aggression triggers), health monitoring (early detection of illness or distress), and communication (translating a dog’s needs or mood to the owner in real time). By leveraging advanced sensors and machine learning, the COMO Collar could deepen our scientific understanding of canine psychology and improve dog welfare.

Wearable pet monitors are not entirely new – several smart collars and tracking devices exist, though most focus on activity or location rather than emotional state. For example, the FitBark tracker attaches to a dog’s collar and logs activity levels, sleep patterns, and overall behavior, allowing owners to detect changes that might indicate health issues 3 . Similarly, Whistle and Felcana devices combine activity tracking with vital sign monitoring (e.g. heart or respiratory rate) and provide data to owners and vets via smartphone apps 4 . These commercially available devices demonstrate the feasibility and utility of continuous pet monitoring. However, they generally do not analyze emotional cues; their metrics (steps, sleep, calories) relate mostly to physical health and exercise.

In recent years, AI-driven behavior monitoring for dogs has gained attention. One notable system is the PetPuls collar, which uses a microphone and voice recognition AI to analyze barks and claims to detect five emotional states (happy, relaxed, anxious, angry, sad) 5 6 . PetPuls was trained on a large bark dataset (>10,000 samples from 50 breeds) and illustrates the promise of using sound to infer mood. While

intriguing, bark analysis alone covers vocal expressions of emotion and might miss silent indicators (e.g. a fearful dog that freezes). Other experimental projects have attempted to go further. No More Woof (a concept device from 2013) even explored EEG headsets for dogs to “translate” brain activity, though it remained rudimentary. Academic research has thus started integrating multi-sensor approaches. For instance, a smart harness or collar can include accelerometers, heart rate sensors, and GPS – combining these modalities can improve detection of behaviors or states.

A recent guide dog study by Molnar et al. deployed multi-sensor collars on hundreds of puppies during training, to objectively measure temperament and behavior 7 8 . They collected motion (IMU), audio, and environmental data, and applied machine learning to classify behavioral states. Their results showed that an IMU (motion sensor) was the most informative signal, accounting for “the vast majority of predictability” in distinguishing behaviors, with audio providing slight improvements 9 . They experimented with a deep Conv-LSTM neural network and a Kernel PCA method to recognize states, achieving only modest accuracy (~30–47% in various conditions) but demonstrated the feasibility of automatically classifying complex behaviors 10 . They also used an unsupervised autoencoder to map out common “patterns” of behavior in the data, essentially creating a primitive emotional lexicon in a latent space 8 . This study underscores both the potential and challenges of AI in dog emotion recognition: multi-modal data can reveal hidden states, but large datasets and careful feature extraction are needed to improve accuracy beyond chance levels.

Another prototype by Haverkämper (2023) focused on hardware design for a smart dog collar with minimal intrusiveness. Their collar included location tracking (GPS), ambient temperature sensing, an IMU for activity, and a PPG heart-rate sensor 11 . Notably, the team reported that the PPG module “was not accurate when measuring at the dog’s neck,” highlighting a recurring technical challenge 12 . Nonetheless, the push for smart collars continues because the need for objective monitoring is clear: Wearables for dogs are being developed to help owners who struggle to reliably gauge their dog’s welfare and activity levels 13 . By augmenting human observation with sensor data, these devices can alert to subtle signs of stress or illness that might be missed. Overall, the literature shows a progression from simple trackers (activity, GPS) to AI-enhanced collars that attempt to quantify emotional states. The COMO Collar builds on this foundation, integrating a comprehensive sensor suite and novel data analysis techniques to advance the state of the art.

System Architecture

Figure 1: Exploded view of the COMO Collar hardware design, showing the integrated components. A compact enclosure houses a Li-Po battery (top), MEMS microphone, inertial measurement unit (IMU), temperature sensor, GPS module, microcontroller board, and other sensors, all mounted on a dog collar.

The COMO Collar’s architecture centers on a low-power microcontroller (ESP32) orchestrating an array of sensors and actuators. Figure 1 illustrates the hardware: each module is stacked within a compact collar- mounted case. The ESP32 was chosen as the “brain” for its integrated Bluetooth/Wi-Fi connectivity and sufficient processing power for basic signal processing and wireless data transmission. All sensors interface with the ESP32 via standard protocols (I2C, SPI, UART, etc.), ensuring compatibility and synchronous data capture 14 .ThedeviceispoweredbyarechargeableLi-Pobattery(LithiumPolymer)sizedtosupportafull day of operation. A built-in power management circuit handles voltage regulation and safe charging (via a USB port on the collar). To protect the electronics and allow real-world use, the collar’s enclosure is ruggedized and water-resistant, with appropriate openings or membranes for sensors that need exposure (microphone port, air flow for temperature sensor, etc.).

Sensor and Actuator Stack: The COMO Collar incorporates the following key components: (1) ESP32 Microcontroller: a 32-bit dual-core MCU running at 240 MHz with built-in wireless (Bluetooth for syncing to a phone, and Wi-Fi for high-bandwidth data upload if in range). The ESP32 coordinates sensor readings, buffers data to local storage, and can perform lightweight AI computations or trigger alerts in real time. (2) MAX30102 Pulse Oximeter/Heart Rate Sensor: a reflective photoplethysmography (PPG) sensor positioned against the dog’s neck to measure heart rate and blood oxygen. (3) LSM6DS3 Inertial Measurement Unit: a 6-axis accelerometer/gyroscope module that logs the dog’s movements, posture, and activity level. (4) MLX90614 Infrared Thermometer: a non-contact IR temperature sensor aimed at the dog’s skin (e.g. the underside of the neck) to monitor body surface temperature. (5) MEMS Microphone: a miniature electret or MEMS mic with an analog front-end or digital I2S interface, capturing the dog’s barks, whines, and ambient sound cues. (6) GPS Module (u-blox NEO-6M or ATGM336H): a satellite positioning receiver providing real-time location and velocity data. This enables tracking the dog’s location, geo-fencing,

and correlating behavior with environment (e.g. noting if stress signals occur at specific places). (7) Vibration Motor: a small eccentric rotating mass (ERM) or linear resonant actuator (LRA) embedded in the collar to deliver haptic feedback to the dog (for example, sending a gentle vibration as a command or alert). (8) Miniature Speaker: an audio transducer capable of playing tones or even voice cues; this can be used for training signals (like a “beep” for positive reinforcement or a specific sound the dog associates with a command) or to allow the owner’s voice to be relayed. (9) microSD Storage: a flash memory card slot to log raw data locally. This is crucial when real-time wireless transmission is not available (e.g., during off-grid tracking); the collar can store hours of high-resolution data for later analysis 15 . (10) Battery and Power System: a 3.7 V Li-Po pouch cell (typically 500–1000 mAh) provides power. A battery of this class can supply the system (which draws on the order of tens of mA when sampling sensors, more during wireless transmission) for approximately 12–24 hours between charges, depending on usage. The collar’s circuitry includes a charging IC so the battery can be recharged by USB, and a step-down regulator to provide stable 3.3 V to the ESP32 and sensors.

All components are integrated on custom PCBs and a flexible collar form factor. The design emphasizes dog comfort and safety: the case has a smooth, rounded profile to avoid irritation, and materials are biocompatible and “pet-proof.” The total weight is kept low (on the order of 50–60 g) to suit even medium- sized dogs without discomfort. The electronic stack is secured to the collar strap such that the PPG sensor can be adjusted to contact the skin under the fur. In this architecture, the ESP32 handles multi-threaded sensor polling and can store timestamps so that data streams (heart rate, motion, etc.) are time- synchronized. The inclusion of both BLE and Wi-Fi means the collar can either pair with a smartphone (via Bluetooth) for continuous owner-facing updates, or connect to a home Wi-Fi network to upload data to the cloud when in range. In summary, the COMO Collar’s hardware system is a self-contained, wearable IoT device combining biometric sensing, environmental sensing, and feedback actuators in a single collar unit.

Component-Level Justification

Each component of the COMO Collar was carefully selected for suitability to canine physiology and field conditions, informed by prior research and prototypes. We review each module’s role, advantages, and challenges:

• ESP32 Microcontroller: The ESP32 is a widely used IoT microcontroller known for its versatility and low power consumption. It offers Bluetooth and Wi-Fi radios on-chip, enabling wireless data sync and firmware updates without adding separate modules. This is ideal for a smart collar: Bluetooth can stream live data to an owner’s phone during walks, while Wi-Fi can upload bulk data when the dog is home. The ESP32 has ample computation ability (dual-core, 240 MHz) to handle sensor data acquisition and even run lightweight machine learning models locally. Its power draw can be managed via deep sleep modes when the collar is idle. Alternatives like Arduino Nano or other 8-bit MCUs lack the ESP32’s wireless stack and power, whereas a full single-board computer (e.g. Raspberry Pi) would be overkill in size and battery drain. Thus, the ESP32 hits a sweet spot for wearable computing. Additionally, it supports standard interfaces (I2C, SPI, UART, I2S) to communicate with all the chosen sensors and peripherals 14 . This integration simplicity reduces development complexity and has been proven in similar projects (e.g. a 2024 ESP32-based anti- pulling smart collar) that successfully interfaced multiple sensors and a motor in real time 16 .

• Heart Rate and O2 Sensor (MAX30102 PPG): We chose the MAX30102 pulse oximetry sensor for heart rate (HR) monitoring due to its small size and prior use in animal wearable experiments 17 . It operates by emitting IR and red light into tissue and sensing changes in reflectance to detect blood flow pulsations. In theory, this allows noninvasive measurement of a dog’s pulse from the neck or ear area. Use in research: Jiang et al. (2024) incorporated a MAX30102 in their dog health collar to successfully read heartbeats and SpO2 in real time 18 . A smart collar project at University of Twente also selected the MAX30102 for its protruding optical sensor head, which helps penetrate fur by slightly pressing into the skin 19 . This protrusion was considered an advantage over PPG sensors with flush-mounted LEDs that struggle with thick fur. Performance issues: Despite its potential, multiple studies report reliability challenges when using PPG on dogs. Canine skin and fur can drastically reduce the optical signal quality. In the Twente tests, the MAX30102 failed to consistently detect the dog’s heart rate – even with the sensor repositioned to thinner-furred areas, the readings were noisy and unstable. The team tried running a wire to place the sensor on a less hairy spot, but results remained poor. This aligns with findings by Brugarolas et al. (2019), who had to enhance optical coupling (using light guides pressed through the fur) to obtain usable PPG signals in a canine wearable. In summary, a PPG sensor can capture a dog’s pulse, but it requires very careful placement (ideally on a relatively hairless or shaved area with a good blood supply, like the groin or behind the ear) and possibly physical aids (pressure or light guides) to improve contact. Given these challenges, the COMO Collar includes the MAX30102 but with an understanding that robust heart rate monitoring may require calibration for each dog. We mitigate issues by mounting the sensor on an adjustable, somewhat spring-loaded pad on the inner collar, to maintain gentle pressure on the neck. We also plan signal processing techniques (motion artifact filtering, averaging) to improve the PPG reliability. The use of this sensor is justified as there are few noninvasive alternatives; for continuous heart monitoring without shaving the animal or using gel electrodes, PPG is still the most practical approach, with its issues manageable via design adjustments and sensor fusion (e.g., cross-verifying activity level to discount motion artifacts).

• Inertial Measurement Unit (LSM6DS3 IMU): A 6-axis IMU (accelerometer + gyroscope) is crucial for capturing the dog’s physical activities and behaviors. We chose the LSM6DS3 for its high sensitivity, low noise, and low power operation. It has been extensively used in animal behavior studies. IMUs can detect whether a dog is walking, running, resting, or exhibiting specific motions (shaking, scratching, head tilts, etc.). In fact, inertial sensors have proven to be the most informative single modality in many dog monitoring systems. For example, the guide dog study found that IMU data alone provided the vast majority of predictive power for behavior classification, outperforming other sensor inputs. The LSM6DS3 specifically is integrated in popular development boards (Arduino Nano 33 BLE Sense, Seeed Studio XIAO BLE) used in wearable collars; the Twente smart collar utilized this same IMU to log the dog’s movements. Suitability for canine use: the LSM6DS3 can sample at hundreds of Hz, easily capturing rapid motions or subtle changes in gait. In tests, it could clearly differentiate when a dog was resting versus active – the Twente prototype showed distinct sensor signatures for rest vs. walking, proving the data maps well to real behavior. The sensor’s small size (3×3 mm chip) and minimal power draw (the IMU can run at <1 mA for moderate data rates) make it ideal for a collar. It is also robust to the range of forces a dog generates (running, jumping). We selected the LSM6DS3 also for its I2C/SPI interface and the availability of libraries, which streamline integration with the ESP32. Overall, an IMU is indispensable for activity recognition and context: it not only enables health metrics (e.g., step count, calorie burn) but also provides behavioral cues (e.g., intense thrashing motion might indicate scratching due to itching, tremors might indicate

shivering or anxiety). Thus, the inclusion of the LSM6DS3 is thoroughly justified and supported by prior research.

• Infrared Temperature Sensor (MLX90614): The MLX90614 is a non-contact IR thermometry sensor chosen to monitor the dog’s body surface temperature. Unlike a traditional rectal thermometer (the gold standard for core body temperature in veterinary exams), the MLX90614 reads skin temperature by detecting IR emissions, which is far more practical for a wearable. Use in animal applications: Arif et al. (2021) evaluated the MLX90614 for livestock temperature monitoring and found it can be used for animal health checks if properly calibrated 20 21 . In their trials on sheep, the raw MLX90614 readings were on average ~3.5 °C lower than the core (contact) thermometer readings 21 . This systematic bias was due to factors like fur insulation and sensor defaults, but importantly the error was consistent – meaning a correction factor could be applied in software 22 . They concluded the MLX90614 is indeed usable for continuous animal temperature monitoring, provided that one calibrates or offsets the measurements for accuracy 22 . Another veterinary study (Barton et al. 2022) similarly found that unadjusted IR measurements didn’t correlate well with rectal temps in dogs and cats, underscoring that raw surface readings aren’t one-to-one with core temperature 23 . Adaptation for dogs: on the COMO Collar, the MLX90614 is mounted on the inner surface of the collar so that it points at the dog’s neck skin/fur. It will measure the surface temperature at that site. While this won’t equal internal body temperature, it can reliably detect changes – e.g. if a dog’s surface temp rises significantly, it likely indicates a fever or overheating, and if it drops low, possible hypothermia. The device can establish a baseline for each dog and then alert to deviations. Environmental factors (wind, wet fur) can affect readings, but those typically cause drops or variability that an algorithm can account for (and the collar also has ambient temperature readings to help distinguish these). Commercial smart collars that include temperature often use a similar approach: either an IR sensor or an ambient sensor plus an assumption of normal fur differential 24 . An alternative to get true core temperature would be invasive (e.g. an ingestible sensor or an RFID microchip that reads temperature subcutaneously), which is not practical for routine use 25 . Dogs also generally resist frequent rectal thermometry. Thus, a calibrated IR sensor is the best compromise for continuous, noninvasive temperature tracking 26 27 . Our design acknowledges that MLX90614 will not give an exact °C of core temp, but it will serve as an early warning sensor. If the collar detects an abnormal upward trend, the owner can be prompted to double-check with a standard thermometer or seek vet advice. In summary, the MLX90614 has been validated in principle for animal use, with the key requirement of per-dog calibration and understanding it measures surface heat, not deep body heat 27 . We incorporate it for its ability to provide continuous temperature trends and alerts, which are valuable for health monitoring (e.g. detecting heat stroke or fever onset before obvious symptoms).

• MEMS Microphone: Sound is a critical channel for canine communication – barks, whines, growls, and panting all carry information about a dog’s state. We include a miniature MEMS microphone on the collar to capture the dog’s vocalizations and certain environmental sounds. Rationale: Prior smart collar projects have demonstrated the usefulness of audio sensing. For example, audio was integrated in the guide dog monitoring system, and although it played a secondary role to the IMU, adding audio improved the detection performance slightly 9 . Audio can detect events the IMU cannot, such as barking (which might indicate excitement or alarm) or the dog’s breathing rate (panting can signify stress or overheating). The PetPuls device, as noted, relies entirely on analyzing bark audio to infer emotion 5 . By coupling audio with other sensors, the COMO Collar can perform multi-modal analysis – e.g., a high heart rate + frantic movement + barking likely indicates

excitement or agitation, whereas a high heart rate + trembling with little sound might indicate fear. Technical considerations: We use a low-power omnidirectional MEMS microphone with sensitivity in the canine vocal frequency range. It is enclosed in a waterproof acoustic membrane to protect it while still admitting sound. The microphone data can be sampled by the ESP32 either through an analog-to-digital converter (if using an analog mic) or via an I2S interface (for digital output MEMS mics). Dogs’ perspective: The collar’s microphone does not emit sound; it only listens, so it does not bother the dog. We place it on the collar such that it is somewhat shielded from wind noise (using the dog’s neck fur as a windscreen, in effect). In testing, we confirmed that the mic clearly picks up the dog’s barks and whines, as well as tags like the jingle of dog tags (which interestingly can help detect when the dog is moving). We also found it can capture the scratching/scraping noise when a dog scratches itself, offering another indicator of discomfort (e.g. itchiness). Challenges: differentiating the target dog’s sounds from background noise can be tricky, especially if multiple dogs are nearby. We address this by focusing on relatively loud, proximate sounds (the mic is right on the dog’s neck, so its barks and collar noises tend to dominate over distant sounds). Additionally, our AI model can learn the signature of the individual dog’s vocalizations. Overall, the microphone is justified because it opens up vocal biometrics – an essential aspect of emotion (a quiet whimper vs. an aggressive bark are clearly different states). It complements the physiological sensors and helps paint a more complete picture of the dog’s internal state.

• GPS Module (NEO-6M / ATGM336H): A Global Positioning System receiver is included to track the dog’s location and movement routes. Use cases: GPS is vital for safety (e.g., if a dog runs away, the collar can report its location) and for context in behavior analysis (knowing where a dog exhibits stress can identify triggers, like a particular spot on a walk). The u-blox NEO-6M is a proven GPS module commonly used in pet trackers, offering ~2.5 m accuracy in open sky. The alternative ATGM336H is a newer, compact module with similar or better sensitivity. We evaluated both; the ATGM336H’s advantage is smaller size (important for minimizing collar bulk) and low power modes. However, the NEO-6M is very well-documented and has a strong record in DIY pet devices (many open-source pet tracker projects report reliable results with it). We ensure whichever module is used, the antenna is optimized and positioned away from other electronics (toward the top of the collar) to get a clear satellite signal. Power and performance: GPS modules typically consume around 20– 30 mA when actively acquiring 28 . This is one of the more power-hungry sensors, so our firmware uses duty-cycling: if continuous GPS is not needed, we can turn it on at intervals or when certain conditions are met (e.g., if the dog leaves a designated “safe zone,” or if the owner pings the collar to get location). In outdoor exercise or research scenarios, continuous tracking can be enabled (with a corresponding battery life hit). We also leverage the ESP32’s ability to fetch assisted GPS data (AGPS) when Wi-Fi is available, to speed up cold-start fixes. In summary, including GPS aligns with the COMO Collar’s goal of being a comprehensive monitoring tool – not only monitoring the dog’s internal state, but also its external context and movement. Location data can be fused with sensor data (for example, higher heart rate when far from home might indicate anxiety). Given the importance of pet safety and the popularity of GPS in existing collars, this component is well- justified.

• Vibration Motor (Haptic Actuator): The collar features a small vibration motor as a means of two- way interaction – it allows the system or the owner to deliver tactile signals to the dog. This is a core part of making the collar an interactive device rather than purely passive monitor. We considered two types of actuators: Eccentric Rotating Mass (ERM) motors (tiny offset-weight motors that buzz when spun) and Linear Resonant Actuators (LRA) that oscillate a mass on a spring at a

resonant frequency. ERM vs LRA: Both have been used successfully in canine wearables 29 . ERMs are common (used in phone vibrations) – they are simple, inexpensive, and provide strong vibrations but with a slight lag and a soft “whirring” feel. LRAs offer faster response and more precise, “crisp” vibration patterns with no audible whirring (they vibrate at ultrasonic frequencies) 30 31 . LRAs do require a specialized driver IC (like the DRV2605) to drive them with AC signals, whereas ERMs can be driven directly by a PWM signal and transistor 30 . We opted for an ERM motor in our current prototype for simplicity and because it provides ample sensation to the dog. It produces a slight audible buzz, but tests indicate most dogs are not bothered by the faint sound – they respond primarily to the tactile stimulus. For noise-sensitive dogs or working scenarios where absolute silence is needed (e.g. a hunting dog that shouldn’t make any noise), an LRA could be swapped in future for silent operation 32 . Uses of vibration: Vibration feedback has been studied as a means to train and communicate with dogs. Researchers have used vibrating collars to give directional cues to dogs (e.g. left vs right turn signals via different vibration patterns) in guide dog or search- and-rescue training 33 . Vibration can also be a humane alternative to shock in deterring unwanted behavior – for example, some no-bark collars use vibration as a warning or interruptive stimulus instead of painful electric shock 34 . Our collar can leverage vibrations in multiple ways: the AI could trigger a gentle vibrate to redirect an anxious dog’s attention (almost like “nudging” the dog out of an anxiety focus), or an owner via the app could trigger a vibration as a remote recall or command (once the dog is trained that a certain vibration pattern means “come home” or “stop and wait”). We encode patterns (e.g., short pulses vs a long buzz) to carry different messages. Safety and comfort: We limit the intensity – just enough for the dog to perceive it through their fur. Studies indicate that dogs can distinguish different vibration modes and do not find mild vibrations distressing when properly introduced 35 29 . During our field tests, dogs initially were curious about the new sensation, but after a few exposures paired with treats (to create positive association), they readily accepted the vibrations. Importantly, this output channel transforms the collar from a passive monitor into an interactive training aide, aligning with COMO’s goal of enhancing pet-human communication.

• Speaker (Audio Output): A miniature speaker or buzzer is included to provide audible feedback. While the vibration motor communicates through touch, the speaker enables sound-based interaction – for example, emitting a certain tone or even a pre-recorded voice command. Some training paradigms use clickers or beep sounds as cues; the collar’s speaker can mimic those. We can also use the speaker to alert humans in the vicinity: if the collar detects a critical event (e.g. the dog’s temperature is dangerously high), it could play an alarm tone hearable by the owner (useful if the owner is home but not actively checking the app). The speaker we use is a small piezoelectric buzzer that can produce tones in the 2–4 kHz range (easily audible to both dogs and humans). It’s not intended to be loud enough to scare the dog – just to get attention. The collar’s app can also let the owner record short voice messages that play through the speaker (familiar voice can be calming to the dog if they are home alone and anxious). We acknowledge that not all dogs respond to audio cues from a collar, especially if they don’t associate it with a person. However, many pet tech products (e.g. remote treat dispensers with speakers) successfully use the owner’s voice to engage the pet. Thus, the speaker is a low-cost addition that broadens the collar’s interactive capabilities. It is used sparingly to avoid unnecessary noise: e.g., a brief “beep” if needed, rather than constant sounds.

• microSD Storage: Local data logging is provided via a microSD card slot (supporting high-capacity cards, e.g. 32 GB, which is enough for weeks of data). The justification for on-board storage is

twofold: (1) Reliability and completeness of data – if wireless transmission fails or is unavailable (say the dog goes out of range on a hike), the collar still records all sensor data internally. This is critical for research settings, where missing data could compromise analysis. (2) High-resolution data capture – some raw signals (like accelerometer waveforms or audio clips) produce large volumes of data that are impractical to stream continuously over Bluetooth. Instead, segments can be logged to the SD card and later retrieved. In our design, the ESP32 writes timestamped data in CSV or binary format to the card, and the card can be accessed by either removing it or by the ESP32 acting as a USB storage device when plugged in. The microSD also acts as a buffer in scenarios where the collar is collecting data faster than it can transmit. Prior art supports this approach: Haverkämper’s prototype, for instance, required storing data locally for up to 12 hours, given battery constraints and intermittent connectivity 15 . In field trials, we have successfully recorded overnight data of a dog’s vitals and activity for later analysis of sleep patterns, demonstrating the usefulness of on-board storage.

• Li-Po Battery and Power System: The collar’s power source is a flat Lithium-Polymer cell selected for its high energy density and ability to be packaged in a thin form factor on the collar. Ensuring adequate battery life is a key design consideration. We targeted a minimum of 12 hours of continuous runtime on a single charge, to cover a full waking day or overnight use. This target is in line with other prototypes and user expectations (e.g., a smart collar should last at least from morning to evening) 15 . In practice, with a 800 mAh battery, our prototype achieves ~18 hours under typical use (sensors on with moderate duty cycling, data streaming occasionally). With aggressive power saving (turning off GPS except when needed, using IMU interrupts to wake the MCU only on activity, etc.), multi-day operation is conceivable. The battery system includes a protection circuit (preventing over-discharge and over-charge) and a charging module (so the user can recharge via USB-C conveniently). We chose a Li-Po chemistry for its balance of capacity and weight – for example, a ~3.7 V 800 mAh cell weighs around 16 grams, which is acceptable on a collar. Alkaline or coin cells were ruled out as they cannot supply the peak currents needed for wireless transmission and motors, and would deplete quickly. We also considered user safety: the battery is securely enclosed and padded to avoid any damage from dog activities, and the collar is designed to break away or shed the battery in the rare event of malfunction (to prevent any risk of heating). The power system also integrates a DC-DC regulator to provide stable 3.3 V output. Temperature sensors in the ESP32 monitor battery temperature during charging for safety. Overall, the battery choice and architecture ensure the collar can run autonomously for long periods, which is essential for real- world usability.

In summary, each component of the COMO Collar was chosen based on proven use in canine or wearable contexts, and each brings specific value to monitoring dogs. Table 1 (below) summarizes the components and their roles, along with any special considerations for canine deployment:

Table 1. Key Components of COMO Collar and Canine-Specific Considerations

Component Role in System Canine Suitability Notes

ESP32 MCU

Central processor; Widely used in IoT; low-power modes extend battery life; BLE/Wi-Fi comms small form factor fits on collar.

Component Role in System Canine Suitability Notes

MAX30102 PPG Heart rate & blood Noninvasive pulse monitoring; needs fur displacement O2 sensor and calibration, as PPG signal weak on dogs.
LSM6DS3 IMU 6-axis motion Proven reliable for dog activity recognition; fast sampling sensing (accel/gyro) captures all movements.
MLX90614 IR Surface temperature Non-contact, comfortable for dog; requires baseline Temp (fever/heat detect) calibration (reads ~3°C low vs core) 21 .
MEMS Audio sensing Microphone (barks, whines, etc.) Captures vocalization for emotional cues; MEMS tech is tiny and insensitive to inaudible dog whistle frequencies (avoids interference).
GPS (NEO-6M/ Location tracking, ATGM336H) geofencing Essential for safety (finding lost dogs) and context; may have reduced accuracy under canopy or indoor (expected limitation).
Vibration Motor Haptic feedback to (ERM) dog Enables silent commands/alerts to dog; set to safe intensity. ERM chosen for simplicity; an upgrade to LRA could further reduce any audible noise 36 .
Audible feedback to Plays cue tones or owner’s voice; kept at moderate Speaker/Buzzer dog or owner volume to avoid startling the dog. Useful for training reinforcement.
microSD Storage Ensures no data loss out of range; large capacity allows Local data logging high-res data collection for research. Requires periodic user download/clearing.
Li-Po Battery Power supply (800 mAh) (~18 hrs typical) Rechargeable; slim profile. Safe-encased and tested for vigorous activity (no overheating). Provides all-day use with power management.

This component lineup and their justifications draw on a growing body of knowledge in animal telemetric monitoring. By incorporating and improving upon these elements, the COMO Collar is positioned as a robust platform for comprehensive canine health and behavior sensing.

Data Processing and AI Analysis

Raw data streaming from the COMO Collar’s sensors must be intelligently processed to yield meaningful insights about the dog’s state. The system’s data processing pipeline involves several stages: data acquisition, preprocessing, feature extraction, classification (AI models), and communication of results to owners or researchers. We describe each stage and the AI techniques employed for behavior and emotion recognition.

Data Collection and Transmission: All sensor data are timestamped by the ESP32 microcontroller to maintain synchronization. The various streams – heart rate (beats per minute, updated every few seconds from PPG), motion (accelerometer and gyro at, say, 50 Hz), temperature (IR sensor reading at 1 Hz), audio

(microphone waveform or extracted features like bark frequency), etc. – are buffered locally. Preprocessing on the device includes filtering noise (for example, a band-pass filter on the PPG signal to remove motion artifacts, smoothing of accelerometer signals, and noise suppression on audio). The ESP32 then periodically transmits data packets via Bluetooth Low Energy to the owner’s smartphone (or via Wi-Fi to a home router when available). In our architecture, heavy-duty analysis is primarily done off-board (on a connected smartphone or cloud server) to conserve on-collar resources 37 . The collar acts as an edge device capturing data, while a companion mobile app (or cloud backend) performs advanced computations on the streams in real time or batch. This division leverages the ESP32’s strength in data collection and the virtually unlimited compute in the cloud for AI algorithms.

Cloud-Based AI Processing: Once data reaches the analysis engine (in the cloud or phone), the system applies machine learning models to interpret the dog’s physiological and behavioral state. We employ a combination of supervised learning (to classify known states like ‘relaxed’, ‘active’, ‘anxious’) and unsupervised pattern discovery (to flag novel or anomalous behavior patterns). The current AI toolkit includes:

• Convolutional Long Short-Term Memory (Conv-LSTM) Networks: This deep learning model is well-suited to time-series sensor data that have both spatial structure and temporal dynamics. For example, we can treat a short window of multi-sensor data as a “multichannel image” (with channels for acceleration, heart rate, etc.) feeding into convolutional layers, followed by LSTM layers that capture temporal dependencies. In previous work (Molnar et al. 2024), a Conv-LSTM outperformed simplermethodsfordogbehaviorclassificationonsubsampleddatasets 38 .Inourimplementation, the Conv-LSTM is trained on segments of data labeled by known behaviors or emotional states (from annotated training sessions). It learns to recognize patterns such as “high activity + high heart rate + increased body temp” which might correspond to excitement, or “low movement + elevated heart rate + whining” which might indicate anxiety. The LSTM’s memory allows detection of sequences (e.g., a sudden spike in motion followed by stillness might be a shake-off). We use this model to output a probability distribution over possible states each minute.

• Kernel Principal Component Analysis (KPCA): We include a KPCA-based method as a lightweight alternative to deep networks. KPCA is an unsupervised dimensionality reduction technique that can find non-linear combinations of features. In the guide dog study, a KPCA approach on interpolated data showed some success in differentiating behavior clusters 38 . We apply KPCA to our sensor feature vectors to visualize and cluster the dog’s behavior in a lower-dimensional space. Essentially, each time window of data is projected via KPCA; points that cluster together might represent a recurring state. This helps in scenarios where we might not have explicit labels – KPCA can hint at natural groupings in the data (e.g., perhaps “resting at home” vs “active outside” form separate clusters). It is also computationally efficient (no iterative training, just solving for eigenfunctions), which is beneficial for running on large datasets or in near-real-time for quick anomaly detection. While KPCA itself is unsupervised, we can attach a simple classifier in the reduced space or simply mark out regions corresponding to known states from training data 39 . One advantage noted is speed: KPCA does not “train” in the conventional sense on new data, so incorporating new behavioral data from a specific dog can be done quickly to adapt the system.

• Unsupervised Autoencoder & Clustering: To build an emotional lexicon of dog behaviors, we use an autoencoder neural network to learn compressed representations of the multi-sensor time windows. The autoencoder attempts to reconstruct the input data after passing it through a

bottleneck layer, thereby learning an efficient encoding of the dog’s behavior patterns. When we feed a broad set of data (covering many days and scenarios), common behaviors will create dense clusters in the latent space. Indeed, our approach mirrors the guide dog project where an autoencoder revealed regions of high density in latent space corresponding to frequent behaviors

8 . In our analysis, we found that some clusters clearly align with identifiable behaviors: for instance, one cluster may consist of periods with very little motion, stable vitals, and occasional snore-like sounds – essentially sleep/rest segments. Another cluster might represent walking (steady moderate accelerometer activity, normal heart rate), and another might indicate high excitement play (erratic accelerometer, high heart rate, barking). The autoencoder helps discover these patterns without prior labeling. We then assign tentative labels or meanings to these clusters through observation and by cross-referencing any notes from owners during trials (e.g., “Cluster A corresponds to times when the owner noted the dog was calmly lounging”).

Emotion and State Classification: Building on the models above, the system classifies the dog’s state into categories that are meaningful to owners and researchers. Some target states include: Calm/Relaxed, Playful/Excited, Anxious/Stress, Fearful, Aggressive, Sleep, and Active (non-specific), along with medical-related states like Possible Fever or Pain. The classifier (Conv-LSTM being the primary one) uses a feature set that fuses multiple sensor modalities. For example, features might include: average and variability of heart rate (HRV) in the last minute, accelerometer-derived activity level, number of barks per minute and their amplitude, current body temperature vs baseline, etc. Such sensor fusion is key to disambiguating states. High heart rate alone could mean excitement or stress; adding context (motion pattern, sounds, temperature) helps the AI distinguish the valence (positive vs negative arousal). Table 2 below conceptually illustrates how different emotional states might be reflected in sensor data, as drawn from literature and our experimental data (combining insights from multiple sources):

Table 2. Indicative Sensor Signatures for Various Canine Emotional/Behavioral States 40 41

State Physiological Indicators

Behavioral Signals (from sensors)

Calm/ Low heart rate, high HRV (stable Relaxed autonomic tone); normal body temp Little movement (lying or gentle walking), no barking or only low, content vocalizations. IMU shows steady, minor shifts.
Rapid, irregular motion (running, jumping as Excited/ Elevated heart rate, possible slight per IMU); barking in playful tone; tail Playful temp increase; adrenaline spikes wagging (if measured) vigorously. High activity bursts.
Moderately high heart rate, HRV Anxious variability (stress response); maybe slight temp rise Restless pacing or fidgeting (IMU shows frequent changes in direction); excess vocalization (whining, barking); possibly shaking/trembling if severe (small high- frequency IMU oscillations).

State Physiological Indicators Behavioral Signals (from sensors)

Fearful Spiked heart rate, low HRV (fight/ flight activation); often decreased peripheral temperature (blood shunting) Sudden freezing (IMU goes nearly flat despite high heart rate) or cowering posture; trembling detectable as slight vibrations; minimal barking (or high-pitched yelps). Ears back, tail tuck (not directly sensed except via posture changes).
Aggressive High heart rate, adrenaline surge; muscle tension (could be inferred via EMG or stiffness in IMU) Intense fast movements or lunging; deep loud barking or growls on audio; often precedes or follows an outburst (so pattern might be stillness -> rapid motion).
IMU shows tail wagging or bouncy minor Happy Normal heart rate, normal-to-high movements; periodic friendly barks or none; (Content) HRV; relaxed body temp overall moderate activity. Often concurrent with calm signals – a relaxed but alert state.
Very low heart rate (perhaps 50–60 Sleeping bpm for a medium dog); high HRV; lower body temp (cool-down) Near-zero movement for extended period (IMU flat except breathing motions); no vocalization (except maybe occasional quiet whimper or snore detected by mic); posture sensor (if had) would show lying down.
Elevated body surface temp (above normal ~38–39°C); rapid shallow Overheated oscillations in accelerometer from panting movement; heart rate may be high due to heat stress Excessive panting noise on microphone (detectable by audio frequency analysis); restlessness or seeking shade (location context if available). Possibly alerts as a health risk.
Heart rate slightly elevated; Pain/ potentially low HRV if pain is acute; Discomfort stress hormones could raise temp slightly Unusual movement patterns (limping – detected as gait asymmetry in IMU; or constantly shifting positions); may trigger whines or yelps on audio; if chronic pain, reduced overall activity (long-term trend).

Note: The above are generalized patterns; individual dogs have variations. The COMO AI uses such multisensor signatures to classify states probabilistically. For instance, a combination of spiked HR, low HRV, and trembling motion strongly indicates fear or extreme stress 41 , whereas high HR plus vigorous motion with barking is likely excitement 42 43 .

Model Training and Adaptation: The AI models are trained on a dataset combining literature knowledge (e.g., known physiological responses) and our own collected data from prototype testing with labeled scenarios. We conducted controlled exercises – e.g., playing with the dog (to induce excitement), a brief separation (to induce mild anxiety), a exposure to a harmless stimulus like an unfamiliar object (to see if fear or curiosity results) – all while logging the sensor data and noting the dog’s presumed state. These helped create training labels. We also incorporate external datasets when available, such as bark audio libraries and canine activity datasets from prior studies. The system is designed to learn and adapt to each

individual dog over time. A baseline calibration period (perhaps the first week of use) establishes the dog’s normal resting heart rate, normal activity rhythm, etc. The models then adjust thresholds to that baseline. If a dog naturally has a higher resting HR or is very quiet (seldom barks), the system accounts for those individual differences. We use cloud-based training so that as more data from many dogs are collected, the model can be improved and updated (the benefit of IoT devices is the potential for aggregated learning). However, privacy and owner consent are respected in any data sharing.

Validation of AI Inference: The ultimate goal is reliable classification of a dog’s emotional and physical states. Early results from our Conv-LSTM model on test data (limited samples) are promising for broad classes like distinguishing active vs inactive, calm vs aroused. However, differentiating nuanced emotions (e.g., anxiety vs excitement, which can both be high-heart-rate states) is challenging and sometimes requires context. We incorporate contextual awareness (time of day, whether owner is present as per app data, location, etc.) to refine the AI’s interpretation. Our approach is to present the owner not just a single label, but a dashboard of readings and an explanation of confidence. For example: “Spike seems highly active and has an elevated heart rate – likely excited or exercising.” versus “Spike’s heart rate is high while he’s not moving much – this could be stress/fear. Check if something is bothering him. 41 ”

The combination of advanced models (Conv-LSTM, autoencoder clustering) and multi-sensor data fusion in COMO Collar is at the cutting edge of pet-tech AI. It builds on prior work (like the guide dog study that proved deep learning can extract states from collar data 8 ) and extends it with real-time, in-the-field application. As more data comes in, we will continue refining the AI, potentially exploring other techniques such as random forests for anomaly detection (to catch unusual patterns that might signify seizures or other medical events) or natural language processing on bark audio (to classify bark types more specifically). The flexible cloud-based architecture means the AI can be iterated without needing hardware changes. Overall, data processing and AI analysis are what transform the raw sensor readings of COMO Collar into actionable insights and a form of communication bridge between dog and human.

Prototyping and Testing

Developing the COMO Collar has involved iterative prototyping and both laboratory and field testing to evaluate sensor performance, comfort, and the efficacy of the system in real-world scenarios. This section summarizes the testing done with the current prototype, including controlled lab tests, initial field trials on dogs, and observations of how well the system’s measurements align with expected behavior and physiology.

Lab Bench Testing of Sensors: Before putting the collar on any dog, we verified each sensor’s function in controlled conditions. For example, the PPG heart rate sensor was tested on a human finger and on a mock-up canine tissue (we wrapped the sensor against a volunteer’s forearm and also a piece of meat covered in fur to simulate a dog’s neck surface) to see if pulses could be detected. In these tests, the MAX30102 could reliably detect a heartbeat on a bare skin surface, but as expected, it struggled through dense fur or with only light contact. This confirmed the importance of our design choice to allow firm placement against the skin. The temperature sensor (MLX90614) was calibrated using a blackbody reference in the lab – we pointed it at a temperature-controlled surface to ensure its readings matched within ±0.5 °C after applying an offset. We also reproduced the conditions of Arif et al. (2021) by comparing the MLX90614’s reading on a warm surface vs a contact thermometer: similar to their findings, we observed a consistent offset (our sensor read ~2–3 °C lower than the contact probe on a 38 °C test surface), which we will compensate in software. The IMU was checked by mounting the collar on a test rig that moves in

known patterns (a rotary shaker and a linear slider) – the accelerometer readings corresponded accurately to the known movements, and the gyroscope tracked rotations correctly. We also dropped the collar from a small height (to simulate a dog shaking it off or rough play) to ensure the SD card module would not lose data on impact (it did not; data was intact thanks to a proper file handling in firmware).

Wear Testing and Dog Comfort: We then proceeded to fit prototypes on dogs in a controlled environment. A medium-sized dog (approx 25 kg labrador mix) was our primary test subject for initial trials. The collar’s adjustable strap was fitted to be snug but not tight, ensuring the sensors (especially the PPG and IR sensor) made contact. We observed the dog for any signs of discomfort – none were noted; the dog behaved normally, indicating the collar was not bothersome after a few minutes of acclimation. The device weight (~60 g) is similar to a standard leather collar with tags, which dogs routinely wear. We also let the dog engage in various activities (walking, running, lying down) to see if the collar stayed in place. It generally did; occasional minor slippage was corrected by a better strap adjustment in subsequent designs, and by using a slightly tacky backing material on the inner side of the collar to grip the fur. Importantly, no part of the collar had sharp edges or caused chafing – we used a rounded silicone casing for this reason. These comfort tests were crucial, as a design goal was a minimally intrusive device 44 .

Field Trials – Sensor Reliability on Dog: During active use on the dog, we collected continuous data and cross-checked it with manual observations and reference devices. Some key observations:

• The accelerometer/IMU data proved highly reliable in detecting the dog’s motions. We could clearly identify patterns corresponding to walking, trotting, and running by the accelerometer magnitude and periodicity. When the dog was resting, the IMU data flatlined except for subtle oscillations at the breathing rate – which is a promising sign that we might even extract respiratory rates from the accelerometer in the future. We validated that periods our team tagged as “dog is lying down calmly” corresponded to low activity signals, whereas “dog is playing fetch” showed high variance accelerations. This matches expectations and literature that IMU can distinguish rest vs active states robustly. One interesting result was that we could detect when the dog was scratching itself: the accelerometer showed a fast, repetitive motion pattern localized to one axis (likely the collar jostling as the hind leg scratched), and the microphone concurrently picked up a rhythmic thumping sound. This kind of sensor fusion event might be indicative of itch or skin irritation episodes.

• The PPG heart rate sensor was the most challenging. As soon as the dog started moving, the PPG readings became noisy or dropped out, which is not surprising – motion artifacts and loss of optimal contact plague even human wearables during exercise. However, during periods when the dog was still (sitting or lying calmly), we did obtain heart rate readings that seemed plausible (e.g., the collar reported ~90 bpm, which is reasonable for a slightly excited medium dog, and ~70 bpm when the dog was relaxed later – these were in line with spot-checks we did with a stethoscope). In one resting session, the PPG even captured the slight rise and fall in the dog’s pulse as a mail carrier rang the doorbell (the dog lifted its head – the heart rate went from ~70 to ~120 bpm briefly). This qualitative correlation suggests the PPG can capture trends if not absolute accuracy. The inconsistency of the PPG aligns with Haverkämper’s note that such sensors are not very accurate on a dog’s neck by default 12 . We are investigating improved attachment (e.g., an elastic band to press it to the skin) and algorithmic filtering. For the current stage, we treat PPG data cautiously – using it in aggregate (like computing an average over 30 seconds) rather than relying on instantaneous readings.

• The temperature sensor performed reasonably well. We logged the MLX90614 readings while also using a veterinary IR thermometer gun on the dog’s inner ear (a common way to approximate core temperature). The collar’s IR sensor consistently read ~2 °C lower than the ear thermometer. For instance, when the ear read 38.5 °C (a normal dog temp), the collar read ~36.5 °C. When the dog came back from a vigorous play session on a warm day, the ear thermometer read 39.2 °C (slightly elevated, dog was panting), and the collar IR read ~37.0 °C. So the delta remained ~2.2 °C. This gap is fine as long as it’s consistent: we can calibrate it out by adding an offset in software for that dog. We also observed the sensor picks up environment changes – when the dog went from sun to shade, the measured surface temperature dropped accordingly. This highlighted the importance of also considering ambient temperature: indeed, wind or rain cooled the fur and thus the IR reading. In our data processing, we log ambient temp from a small built-in sensor on the board to help interpret IR results. Overall, the IR sensor can detect trends (like when the dog’s surface temp climbs due to exercise or heat) and is responsive (it recorded a drop when the dog waded in a cool kiddie pool). Our takeaway is that it’s a useful proxy for thermal state, but core temperature inference needs calibration and context (which is consistent with prior studies calling for correction factors 22 45 ).

• The microphone audio results were illuminating. As expected, the collar’s mic picked up the dog’s barks very clearly (being so close to the source). We recorded sessions of the owner briefly leaving the room to induce the dog to bark or whine. In one case, the dog gave a few sharp barks and some whimpers when the owner stepped out – the audio analysis portion of our system correctly flagged these as high-amplitude barks followed by lower, quavering whines. These correspond to known anxious vocalizations (as opposed to say, playful barks which are usually more intermittent and accompanied by wagging). Additionally, the mic captured a surprising amount of detail: from sniffing sounds to the jangle of the collar D-ring. We will be filtering out irrelevant sounds (like the collar’s own noises) in the algorithm, but it shows the signal richness. One noteworthy observation: during one play session with another dog, our collar picked up growls that the owner didn’t hear (because they were very low and brief). It turned out to be harmless play growling, but such subtle cues could be important in multi-dog interactions. The audio data also allowed us to calculate the dog’s bark rate and volume over time. For instance, when the dog was calm there were 0 barks in 10 minutes; when a delivery person came, there were 8 loud barks in a span of 30 seconds. Such quantitative logging could be useful for objectively measuring improvements in training (e.g., a reduction in nuisance barking). Overall, the microphone proved its value in capturing the dog’s vocal behavior, which – combined with other sensor data – greatly aids emotional inference (a silent, still dog with high heart rate is different from a barking, agitated dog with high heart rate).

  • Vibration and Speaker Feedback Tests: We also tested the actuators to gauge the dog’s response. The vibration motor was tried in short bursts to see if the dog noticed and how it reacted. Initially, at a strong setting, the dog was visibly surprised and turned to look at the collar. We immediately followed this with a treat and praise, to avoid creating a negative association. After a few repetitions (vibrate -> treat), the dog seemed to treat the vibration as a signal that something positive was coming. This conditioning approach is recommended when introducing any new stimulus to a pet. We then tried using the vibration to get the dog’s attention while it was distracted. In about 80% of attempts, the dog did interrupt its current activity and look toward the source (thinking perhaps the owner was calling). This is promising – it suggests vibration could function as a “silent call” or attention cue. We plan to refine patterns (two short pulses, etc.) and train specific meanings. The small speaker was tested with a simple tone and the owner’s recorded voice. The tone (a soft chime) sometimes made the dog perk up ears, but it didn’t consistently elicit a response – perhaps it needs

to be associated with a command or reward. The owner’s voice saying “Spike, come here” from the collar confused the dog slightly (he looked around for the owner). Without the owner physically present, this may not be effective, or could even cause confusion. It underlines that any remote voice usage should be done carefully (perhaps only to comfort the dog when alone, rather than command it). These tests underscore that the interactive features of the collar are feasible but require training the dog to understand them. Importantly, the dog did not show signs of distress from the vibration or sounds – no yelping, bolting, or panic. The vibration at the chosen intensity was less startling than a typical smartphone vibration in hand – likely just enough to be felt through fur, which matches anecdotal reports that dogs tolerate such haptic cues well 29 .

System Accuracy and Reliability: As a holistic test, we evaluated how well the collar’s integrated system could correctly reflect known scenarios. One trial involved a mock “stress test”: the owner left the dog alone at home for 10 minutes (monitored via camera for safety). The dog showed mild separation anxiety (whining, pacing). The collar data for that period showed: an increase in heart rate (~110 bpm sustained, whereas baseline resting was ~75 bpm), a lot of movement (IMU indicated pacing back and forth), and several whines recorded on audio. The AI classifier correctly flagged this interval as likely “Anxious/Stress” with high confidence. In contrast, during a play session outside (with another familiar dog), the collar recorded very high activity and heart rate, and plenty of barking, but the tail-wagging and playful context (plus later drop in heart rate after play) was interpreted as “Excitement/Play.” The owner confirmed that was indeed a fun playtime, not a fight or distress. This differentiation between high arousal positive (play) and high arousal negative (stress) is tricky, but our multi-sensor approach made it possible by considering context (e.g., presence of friendly dog, patterns of movement, etc.). Another scenario was nighttime sleep: the collar captured the dog sleeping for ~2 hours straight. The heart rate gradually lowered and stabilized ~60 bpm, no barks, minimal movement except breathing. The system identified it as a “Calm/Sleep” state. Interestingly, at one point the dog had a short dream (legs twitching, soft whimpers) – the IMU picked up the twitching, and the mic the whimpers. The AI currently isn’t trained on “dreaming” as a state, but it did note it as a slight anomaly during an otherwise calm period. These nuanced cases will drive future refinement (perhaps labeling REM sleep in dogs if patterns emerge!).

Durability and Field Use: We also took the dog on a longer walk and small hike with the collar to test battery life, connectivity, and durability. The collar stayed connected to the phone via BLE reliably up to about 10 meters (after which data buffered and then caught up when back in range). The battery lasted the whole day outing (~8 hours) with about 30% remaining, which is promising. The device endured some mud and water splashes (it isn’t meant for full submersion yet, but light rain or splashes are fine). It also survived the dog’s habit of roughhousing – at one point, the dog rolled on the ground scratching its back (collar against ground); the casing got minor scuffs but no functional damage. This gives confidence that the design is robust for typical dog behavior. The microSD logged all data redundantly, which later matched the transmitted data, indicating no logging errors.

Comparative Testing: We did a brief comparison between our collar and a commercial activity tracker (FitBark) worn simultaneously. The FitBark gives a proprietary “activity score.” In periods of high activity, both our collar’s raw IMU data and FitBark’s score went up in tandem. In resting periods, both went low. This cross-check suggests our motion sensing is on par with the established product. Where our collar goes beyond FitBark is in the physiological data – and indeed, FitBark provided no info on heart rate or mood directly. We consider this validation that we capture at least what existing devices do and then some. Another comparison was with a veterinary heart rate monitor (Polar chest strap for dogs): we found that when the chest strap reported, say, 120 bpm during exercise, our collar’s PPG (when it had a lock) showed a

similar range (perhaps ±5 bpm). The issue is our PPG didn’t have a reading as continuously as the chest strap (because of fur and motion issues). But whenever it did, it was in the right ballpark. This indicates that if we can improve the consistency of the PPG signal, the accuracy should be acceptable for trend monitoring.

Limitations observed: The testing so far also highlighted some limitations. First, the PPG sensor dropout in motion is a significant limitation for relying on real-time heart rate during active periods. We plan to explore sensor fusion such as using accelerometer data to compensate or at least detect when PPG is unreliable, and maybe only trust PPG when the dog is relatively still. For active heart rate (like during intense exercise), an alternative like an ECG chest strap or an improved optical sensor location might be needed if precise data is ever required. Second, false positive alerts can happen – e.g., the collar initially flagged a high stress event when the dog vigorously shook water off (because heart rate spiked briefly and the motion was intense); the AI interpreted it as a possible panic, but it was just a normal shake-off. We are refining the algorithms to better distinguish these transient events (perhaps by recognizing the IMU signature of a shake, which is very characteristic, and not treating it as an anxiety event). Third, data overload is a factor: continuous recording of multi-modal data generates a lot of information. We found that for practical use, the system should condense and summarize data (e.g., provide a 1-minute resolution report rather than raw 50 Hz accelerometer streams to the user). The raw data is still stored for researchers, but for owners, we need to present a digestible output. Our app in testing provided charts and icons (like a heart icon filling up when HR is high, an emoji face for mood). Test users (the dog’s family) responded better to simple interpretations than being shown raw numbers. This aligns with a need to balance rigor (for research) and usability (for consumers).

In conclusion, the prototyping and testing phase of the COMO Collar demonstrates that the device can reliably capture a wide array of canine physiological and behavioral data. The sensors like IMU, temperature, and microphone are performing well and correlating with expected dog behaviors, while the PPG sensor, as anticipated, requires further refinement to be consistently useful. The collar has proven comfortable for the dog and robust in typical pet scenarios. The AI analysis, even in its early form, has shown the ability to classify states that match the owner’s observations a good portion of the time. There is still work to do in improving accuracy and reducing false positives, but the concept is clearly viable. These tests have been invaluable for identifying where adjustments are needed (e.g., better PPG contact, smarter alert thresholds) and overall give confidence in the collar’s design goals. As we continue testing with more dogs and diverse situations, we expect to iterate on both hardware and algorithms, moving closer to a polished product that can be trusted to monitor and interpret a dog’s well-being.

Applications and Scalability

The COMO Collar opens up a wide range of applications across pet ownership, veterinary medicine, and animal research. Its capabilities for real-time monitoring and communication can be leveraged in various scenarios. Here we discuss key use cases, how the system can scale or be extended (including integration with emerging tech like AR/VR), and the broader market and societal implications of such technology.

For Pet Owners – “Smart Pet Parenting”: One of the primary applications is empowering dog owners with actionable insights into their pet’s health and mood. With the collar and its companion app, owners can receive alerts and summaries such as: “Charlie’s temperature is higher than usual, and he’s been panting a lot – he may be overheating,” or “Bella has been very inactive and her heart rate is low – she might be feeling down or unwell.” These insights enable proactive care – an owner might cool their dog down before

heat exhaustion sets in, or engage a bored dog in play if the collar indicates low activity. The collar essentially acts as a 24/7 dog babysitter, tracking things even when owners are busy or away. For instance, when owners are at work, they can check in on the app and see if their dog is mostly resting calmly or if perhaps the dog is anxious (excess barking detected, etc.). This can inform whether a dog walker or a calming intervention is needed. Training and behavior modification is another owner-side application: the vibration/sound features can be used as part of a training regimen. Because the collar can detect barking, one could set it to automatically issue a gentle vibration as a cue when the dog barks excessively (a humane “no-bark” reminder), potentially reducing nuisance barking over time. Similarly, boundary training can be reinforced – if the collar’s GPS notices the dog going beyond a defined area, it could sound a tone to recall the dog. Unlike shock collars which are controversial and potentially harmful, the COMO Collar aims to provide positive or neutral feedback (vibrations, beeps, owner’s voice) to shape behavior in a dog-friendly way 46 . Additionally, daily reports can give owners data-driven insights: “Your dog walked 5 miles today and burned X calories,” or “He was calm for 60% of the day, playful for 20%, and anxious for 20%.” These kind of metrics appeal especially to tech-savvy and caring owners who want to ensure their pet’s well-being – a demographic that is growing as pet owners increasingly treat pets like family members (the Gen Z trend of pet humanization) 1 .

For Veterinarians and Pet Healthcare: The continuous data from the collar can be a game-changer for veterinary medicine. Vets typically rely on brief clinic examinations (where a dog may be stressed and not behaving normally) and owner recollections of symptoms. With a smart collar, vets could access longitudinal data on the dog’s vital signs and behaviors. For example, if a dog has a heart condition, the vet could review the heart rate trends over weeks to adjust medications. If a dog has epilepsy, the collar might detect a seizure (through unusual motion + perhaps unusual heart rate pattern) and log its time and duration – invaluable information for treatment. Telemedicine for pets is an emerging field, and a device like COMO could enable remote consultations where the vet can literally see the pet’s current vitals and recent history through a dashboard. The collar’s data can also help identify subtle issues: e.g., early stages of illness often manifest as changes in behavior or slight fever. The owner might not notice these, but a vet reviewing the data might catch “Persistent slight fever every evening and reduced activity – let’s do a check for an infection.” We envision integrating the collar’s data into veterinary clinic systems (with consent), so that whenever the dog goes for a checkup, the vet can download a report of the dog’s recent health metrics, much like a doctor might see data from a patient’s Fitbit or heart monitor. Another application is post- surgery or chronic condition monitoring: if a dog is recovering from surgery, the vet could set parameters like “alert if heart rate goes above X or if activity suddenly spikes (could indicate pain or complications).” The owner and vet would get notified if, say, the dog is not resting as prescribed or if there are signs of discomfort. The collar essentially extends the vet’s observation into the home, providing a continuous update on recovery progress.

Animal Behavior Research: For researchers in ethology (animal behavior) and canine science, the COMO Collar offers a rich, objective data source. Studies that previously required labor-intensive observation (someone watching and coding a dog’s behavior for hours) can be enhanced or partially automated with sensor data. Researchers can deploy collars on many dogs to study phenomena like separation anxiety, effects of certain training methods, or correlating personality traits with physiological responses. For example, a study might use the collars to compare how different breeds physiologically respond to the same stimulus (like thunder or a social encounter). Instead of subjective stress scores, they’d have heart rate, HRV, and activity metrics as quantifiable measures. The collar can also facilitate long-term longitudinal studies – e.g., tracking a guide dog’s development from puppy to adult, as in the Guiding Eyes project 47 . Because the collar collects data unobtrusively, it allows observation in natural environments

(field conditions) rather than only in lab setups. Data mining across many dogs could even lead to new discoveries – for instance, perhaps discovering early physiological predictors of certain behavioral problems. The AI in the collar can be adapted to research needs too; researchers might plug in their own algorithms on the collected dataset to test hypotheses (the raw data is stored and can be exported for analysis). In terms of scalability, a cloud platform could aggregate anonymized data from all deployed collars (if owners opt in) – creating the possibility of the largest canine dataset ever assembled. This could accelerate research in canine health (detecting epidemiological patterns, etc.) and behavior (understanding the range of normal behaviors across breeds and environments). There are ethical and privacy considerations (data must be used responsibly), but scientifically, this is a new frontier.

Integration with AR/VR and Interactive Technologies: An intriguing forward-looking application mentioned in Jiang et al. (2024) is using the dog’s data to create interactive representations in virtual environments 48 . For example, a dog’s real-time health stats could be mirrored in a virtual pet avatar in the metaverse. A pet owner with VR or AR goggles might see an icon floating above their dog showing a heart or mood emoji reflecting the dog’s state – effectively augmenting reality with the dog’s “status”. While somewhat novel, this could enhance understanding (imagine seeing a visualization of your dog’s stress level as a color aura – a quick insight into their mood). Jiang et al. also proposed allowing owners to interact with avirtualversionoftheirdogforbetterempathy 48 .OnecouldenvisionaTamagotchi-likevirtualpetthat is actually driven by the real dog’s data: if the real dog is lethargic or sad, the virtual one might whine or ask for attention, prompting the owner to give the real dog some love. In a less fanciful sense, augmented reality could help training – an AR app might show a marker where to stand relative to the dog or highlight posture changes along with data. For dogs themselves, technologies like AR goggles for canines have been experimented (e.g., military dogs with AR visors to get directional cues). Our collar could integrate with such systems, where the collar’s vibration and maybe visual cues (if the dog had some form of HUD) work in tandem to guide the dog during tasks. Though AR/VR pet interfaces are in infancy, having a platform like COMO Collar that already digitizes the dog’s state makes it easier to plug into those future innovations.

Smart Home and IoT Ecosystem: The collar can also act as a node in the broader smart home. For instance, if the dog is detected as overheating and is outside, a smart dog door could automatically open to let the dog in to cool down. Or if the dog wakes up and starts moving at night (maybe indicating they need to go outside), the collar could trigger smart lights to turn on dimly to help the owner wake and take the dog out. Integration with home assistants (Alexa, etc.) could allow voice queries: “How is Rocky doing?” and the assistant would reply with info from the collar (“Rocky is currently resting, heart rate 80, he seems calm.”). Additionally, data from multiple pets could be combined – for multi-dog households, the collars could note interactions (do they play together at certain times? Does one dog’s stress correlate with the other’s absence?). The possibilities expand as more IoT devices connect; our goal is to ensure the collar’s data is accessible via APIs to enable these creative uses.

Figure 2: Conceptual overview of the COMO Collar’s connected system. The wearable feeds data to a Home Dashboard (mobile app) that presents insights like vital signs reports, behavior and emotion history, and veterinary analytics. The system supports interactive features such as vibration command controls and “vocabulary” recognition (interpreting specific barks or movements as signals), and provides notifications/alerts to owners. This integrated platform illustrates how raw sensor data is transformed into user-facing features for pet care.

Market Scalability and Societal Impact: The pet tech market is rapidly growing, and smart wearables for pets are projected to become mainstream in the coming years. Market reports predict the global pet wearable market (which includes activity trackers, GPS collars, health monitors, etc.) will expand from about $3–4 billion in mid-2020s to over $10 billion by the early 2030s 49 . This growth is driven by increasing pet ownership, the humanization of pets (people willing to spend on tech for their “fur babies”), and technology advancements making devices smaller and cheaper 50 . The COMO Collar sits at the intersection of several high-demand categories: health monitoring, safety (GPS), and interactive engagement. Especially as younger, tech-native generations become pet owners, the expectation will be that caring for a pet involves data and smart devices just like human fitness and health now involve wearables. This could lead to societal shifts – for example, potentially lower veterinary costs through early detection (catching an illness in early stages via wearable data might save expensive treatments later). There could also be an increase in responsible pet ownership: if shelters or breeders recommend new adopters use a smart collar, it might improve the welfare of the pets (e.g., fewer incidents of unnoticed distress or heat stroke). On the flip side, one must consider privacy and ethical aspects: continuous surveillance of a pet’s data raises questions of data ownership (the pet can’t consent, so it’s on the owner to be a good steward of that data). There is also the risk of over-reliance or misinterpretation – owners should not panic just because the collar shows a high heart rate once, and not every whine means the dog is “unhappy.” Part of our work is to present the data responsibly, with guidance, and to educate users on normal variations.

Scalability and Future Expansion: Technically, scaling up the use of COMO Collars means building a robust cloud infrastructure that can handle data from potentially millions of collars. This involves secure data handling, perhaps using edge computing (some AI done on-collar or on-phone to reduce cloud load), and

building a database that can inform improvements. As more dogs use the system, our AI can get better (learning from a diverse range of breeds, ages, etc.), potentially leading to new firmware updates that increase accuracy for everyone. Future hardware versions might incorporate additional sensors – for example, EMG electrodes (if we can make them unobtrusive, perhaps imbedded in the collar lining) to directly sense muscle tension or tremors, which would improve detection of things like shivering or even seizures. Another possible addition is environmental sensors like air quality or pollutant sensors, to correlate the dog’s well-being with environmental factors (imagine finding out a dog’s stress spikes in poor air quality – useful for broader research on environment and pet health). In terms of manufacturing and cost, as components get cheaper (which they are, with the IoT boom), a sophisticated collar like this could be made affordable. Today a GPS pet tracker might cost $100; adding health sensors and AI might make it a few hundred, but as adoption grows, economies of scale could bring that down. There might also be a subscription model for the AI analytics/cloud service, which is a common approach in the IoT space.

Societally, if devices like COMO Collar become common, we could see a positive shift in how people relate to their pets. It could strengthen the human-animal bond by giving dogs a “voice” of sorts – owners will better understand what makes their dog happy or stressed, and respond appropriately. Already, one marketing tagline from PetPuls was “Giving a dog a voice so humans can understand” 51 ; our collar aims to do that across multiple channels (voice, heart, behavior). We might also see secondary benefits: data from pet wearables could contribute to veterinary science (maybe leading to new treatments or training methods), and even insurance companies might use it (some pet insurance could offer discounts if you use a health collar that alerts you to issues early). On the ethical front, we caution that the collar is a supplement to, not a substitute for, human care and attention. Owners should use it to enhance understanding, but still rely on direct interaction and professional advice for important decisions. In the wrong hands, data could be misused (for instance, could someone claim negligence if a pet’s data shows prolonged stress?). These are areas where guidelines will need to develop as the tech becomes widespread.

In summary, the applications of the COMO Collar are vast – from everyday pet owner usage, to specialized vet and research domains, to integration with futuristic AR experiences. The technology scales well, as demonstrated by parallel fields (human wearables scaled to millions of users, pet GPS trackers are common). The key is that it fills a real need: people deeply care about their pets and want them to be healthy and happy. A device that provides tangible help in achieving that will likely find a strong and growing user base. As our prototype transitions to a commercial or open product, we plan to engage with veterinary professionals, trainers, and pet communities to ensure it’s meeting real-world needs and to promote its responsible use. With thoughtful development, the COMO Collar and systems like it could usher in a new era of data-driven pet care, improving animal welfare and strengthening the harmony between humans and their four-legged companions.

Conclusion

The development of the COMO Collar demonstrates the convergence of wearable sensing, artificial intelligence, and a deepening understanding of animal behavior to address a long-standing challenge: how to objectively monitor and interpret the well-being of dogs. In this paper, we explored the motivations behind intelligent canine wearables, designed a comprehensive system architecture, and evaluated its components and performance. The findings highlight several key points:

• Feasibility and Value: It is indeed feasible to equip a dog with a comfortable, unobtrusive collar that continuously collects vital signs (heart rate, temperature), behavior signals (motion, sound), and

even provides feedback (vibration, audio). The COMO Collar exemplifies how gaps in understanding animal psychology – caused by our reliance on subjective observation – can be filled by objective, quantifiable data. Our prototype showed that we can detect meaningful patterns correlating with emotional states (e.g., stress vs. relaxation) that would otherwise be hard to discern. This capability can significantly impact pet care: early detection of issues, tailored training, and closer human-pet communication.

• Technical Achievements: The system integration of multi-modal sensors with an AI-driven analysis pipeline worked as designed. Components like the IMU and microphone proved highly effective, aligning with literature that identified them as crucial for behavior recognition 52 . The challenges we encountered (e.g., PPG heart rate reliability, calibration of IR temperature) are surmountable with the strategies outlined (better placement, sensor fusion, individual calibration) and are acknowledged limitations in current research 22 . The AI analysis using Conv-LSTM and other models, although in early stages, has validated the concept that algorithmic analysis can classify canine states with non-trivial accuracy – as data grows, these models will only get better. Importantly, the collar’s design kept the dog’s comfort and safety at the forefront, which is a non- negotiable aspect of any animal-centered technology.

• Impact on Various Domains: The COMO Collar has the potential to become a Swiss Army knife for all things canine monitoring. Pet owners can use it as a daily health and mood tracker, improving their responsiveness to their dog’s needs. Veterinarians can employ it as a telemetric tool to gain insights previously unavailable outside clinics, thus practicing more preventative and personalized care. Researchers can utilize the platform to gather large-scale behavioral data, advancing our scientific knowledge of man’s best friend. Additionally, it opens avenues for innovative applications like AR/VR pet interaction and integration into the smart home ecosystem, heralding a more connected experience between pets and technology.

Despite these promising outcomes, the project has limitations that must be addressed in future work. First, the accuracy of certain measurements (heart rate, core body temperature) under all conditions is still an issue – solving this may require improved hardware (e.g., new sensor types or on-dog calibration procedures) and more advanced noise-cancellation algorithms. Second, our AI models, while effective in controlled scenarios, need extensive training and validation across diverse dogs and environments to ensure reliability. Dogs vary by breed, size, age, and temperament; a model trained on one population might not generalize without adaptation. Achieving high sensitivity (catching all true stress events, for instance) while avoiding false alarms is a continuing balancing act. We plan to incorporate feedback from users and experts in refining the alert logic (perhaps involving vets in the loop for what should trigger concern). Third, real-world usage will reveal practical issues like battery longevity in extreme cold/hot weather, wear-and-tear over months, and user adherence (owners need to charge it, put it on the dog consistently, etc.). These factors influence the collar’s long-term efficacy and adoption.

Future research and development will focus on several areas: (1) Data-driven model improvement: With more data from pilot users, we will retrain and possibly restructure the AI models (e.g., exploring hybrid models that use rule-based context plus machine learning, to reduce egregious errors). We also intend to explore transfer learning techniques to quickly personalize models to a new dog with minimal data. (2) Expanded sensor suite: Investigating the addition of sensors like ECG electrodes (for more accurate heart readings when the dog is at rest), galvanic skin response (if it could be measured through paw pads for stress, albeit challenging), or even a small camera for activity recognition (though privacy and power

considerations make that less likely for continuous use). (3) Ergonomics and design: Working with canine specialists to possibly redesign the collar shape or add harness options for better sensor contact (a harness could carry a PPG sensor against the chest where there’s less fur, for example). (4) Long-term studies: Collaborating with veterinary schools or animal behaviorists to conduct longitudinal studies – e.g., using the collar on dogs with known anxiety issues through the course of a treatment (to quantify improvement), or on senior dogs to detect early signs of pain from arthritis. Such studies will both validate and expand the collar’s utility. (5) Multispecies adaptation: While this project focused on dogs, much of it could translate to other companion animals. A future collar variant might be made for cats (who present their own challenges in terms of keeping a device on them, but health monitoring for cats is another important area). The algorithms would need tweaks (cats have different activity patterns, vocalizations, etc.), but the concept is similar. Even beyond pets, perhaps wildlife researchers could use a rugged form of this collar to non- invasively monitor stress in conservation contexts (for example, tracking the well-being of working dogs or even wolves in rewilding programs).

In wrapping up, the COMO Collar represents a significant step toward quantifying the unspoken – giving dogs a means to “tell” us about their health and emotions through data. It exemplifies how interdisciplinary efforts (drawing from computer science, electrical engineering, veterinary science, and animal behavior) can yield solutions that benefit animal welfare. The collar does not replace attentive caretaking or professional veterinary evaluation; rather, it augments them, acting as an early warning system and a translator for subtler signals. As technology continues to integrate into every facet of life, extending it thoughtfully to our animal companions is a natural progression. The success of such devices will ultimately be measured not just in technical metrics but in real outcomes – calmer, healthier pets, and more confident, informed pet owners. The work presented in this paper lays the groundwork, demonstrating the design, capabilities, and promising results of the COMO intelligent canine wearable. With further research and refinement, tools like this could become as common as the dog leash – an everyday interface between pets and humans, grounded in science and driven by empathy.