💻 computer science

ActiveVital: Geometry-Aware Embodied Vital Signs Monitoring for Home Healthcare Robots

ActiveVital is a vision-guided framework for home healthcare robots that actively adjusts sensing geometry to align radar with the human chest, thereby significantly improving the accuracy of contactless respiration and heart rate monitoring in unconstrained environments.

Original authors: Yuxuan Hu, Shihao Li, Yang Xiao, Gen Li, Feng Xu, Jianfei Yang

Published 2026-06-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yuxuan Hu, Shihao Li, Yang Xiao, Gen Li, Feng Xu, Jianfei Yang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a home robot designed to be your companion, helping you stay safe and healthy while you relax, sleep, or move around your living room. To do this effectively, the robot needs to know your vital signs—specifically, how fast you are breathing and your heart rate—without ever having to touch you.

The paper introduces a new system called ActiveVital that solves a tricky problem: how to "listen" to your heartbeat and breathing using radar when you aren't standing perfectly still or facing the robot.

Here is how it works, broken down into simple concepts:

The Problem: The "Flashlight" Effect

Think of a millimeter-wave radar (the robot's "ears") like a flashlight beam. It can only "see" motion that is moving directly toward or away from it (like a car driving straight at you). It cannot easily detect motion that is moving sideways.

Your Heartbeat: Your heart beats are tiny movements on your chest. If the robot is looking at you from the side, your chest is moving mostly sideways relative to the robot. The radar beam misses most of that motion, making the heartbeat signal very weak or invisible.
The Old Way: Most robots just stand still and hope you stay perfectly aligned with them. If you shift in your chair or turn your head, the "flashlight" misses the target, and the data becomes noisy or wrong.

The Solution: The "Dancing Robot"

ActiveVital changes the game. Instead of asking you to stay still, the robot moves itself to find the perfect angle. It treats the angle of its radar as a control knob that it can turn.

The Eyes (Vision): The robot uses a camera to look at you and find your skeleton. It specifically looks for your shoulders and hips to calculate exactly where your chest is, like a painter finding the center of a canvas.
The Brain (Alignment): The robot compares where your chest is to where its radar is pointing. If your chest is slightly to the left, the robot's arm swivels left. If you are slightly up, it swivels up.
The Goal (The "Normal" Angle): The robot keeps adjusting until its radar beam hits your chest almost perfectly straight on (like a flashlight hitting a wall head-on). This maximizes the "radial" motion the radar can see, making the heartbeat signal loud and clear.

The "Noise-Canceling" Headphones

Even with perfect alignment, you might still wiggle a little. To handle this, ActiveVital uses a special signal processing trick. Imagine listening to a quiet song in a noisy room; you need a filter to block out the background chatter. The system uses a "differential phase enhancement" module to filter out random noise and isolate the rhythmic patterns of your breathing and heartbeat, even if you shift slightly.

What They Found

The researchers tested this system in three different scenarios:

Static Sensing: The robot sits still, and the human tries to stay perfectly aligned (like holding a pose).
Proximity Sensing: The robot moves closer to you but doesn't adjust its angle (like leaning in to hear a whisper but still looking at the side of your face).
ActiveVital (Geometry-Aware): The robot moves and angles itself to hit your chest perfectly.

The Results:

Breathing: ActiveVital was just as accurate as the "perfect pose" method, but it worked even when you were moving freely.
Heartbeat: This is where the magic happened. In the "Proximity" method (just getting close), the heartbeat error was huge (about 13 beats per minute off). In ActiveVital, the error dropped to just 2 beats per minute.
The Takeaway: By actively steering the radar to face your chest directly, the robot can monitor your heart rate as accurately as if you were holding a perfect pose, but without you having to do anything.

In a Nutshell

ActiveVital is like a smart robot that doesn't just stand there waiting for you to cooperate. Instead, it actively "looks" for your chest, "steers" its sensors to get the best possible view, and "cleans up" the signal to give you a reliable health check, all while you go about your normal day. It turns the robot's movement into a superpower for sensing.

Technical Summary: ActiveVital

Problem Statement

Home robots require reliable, non-contact vital signs monitoring to ensure user safety and enable health-aware interaction. While millimeter-wave (mmWave) radar is a promising solution due to its sensitivity to sub-millimeter thoracic motions, its effectiveness is fundamentally constrained by observation geometry. mmWave radar only measures the radial component of motion along the line of sight. In unconstrained home environments, arbitrary robot–human orientations often result in angular misalignment, which attenuates the measurable signal—particularly for heartbeat components with extremely small amplitudes.

Existing approaches suffer from two primary limitations:

Static Sensing: Relies on users to manually maintain alignment with a fixed sensor, which is impractical for long-term, unobtrusive monitoring.
Proximity Embodied Sensing: Mobile robots reduce the distance to the subject to improve signal strength but fail to regulate the anatomical alignment (observation angle). Consequently, these methods lack a closed-loop mechanism to optimize vital sign observability, leading to unstable heartbeat estimation when the robot and human are not perfectly aligned.

Methodology: ActiveVital

The authors propose ActiveVital, a vision-guided embodied sensing framework that treats sensing geometry as an explicit control variable. The system reformulates vital signs monitoring from passive signal recovery to active geometric regulation. The pipeline consists of three stages:

1. Vision-Guided Chest Localization

To establish a geometric reference without user cooperation, the system employs ViTPose to detect human skeletal keypoints (shoulders and hips).

Chest Reference Construction: A chest anchor point ( $P_{ref}$ ) is interpolated along the torso axis between the shoulder and hip midpoints.
Anatomical Bias: The vertical coordinate is biased toward the upper chest ( $\alpha = 0.75$ ), and a horizontal offset is applied ( $\beta = 0.1$ ) to shift the reference toward the cardiac-adjacent region, compensating for lateral anatomical asymmetry.

2. Closed-Loop Pose Regulation

The robot actively adjusts its sensor pose to align the radar line-of-sight with the chest reference.

Error-Driven Control: Pixel deviations between the radar's optical axis and the chest reference are converted into motion commands.
Spatial Mapping: Using depth information and camera intrinsics, pixel errors are mapped to metric displacements ( $\Delta L, \Delta H, \Delta D$ ).
Execution: The robotic arm steers the radar toward near-normal incidence relative to the thoracic surface. Once aligned, the robot holds a fixed pose for a short window to acquire data, ensuring maximal radial observability ( $\theta \to 0$ ).

3. Geometry-Constrained Phase Processing

To ensure stable signal extraction under residual pose variations, the system employs a differential phase enhancement module:

Phase Unwrapping & Differencing: Raw phase sequences are unwrapped and subjected to first-order temporal differencing to suppress low-frequency drift and enhance periodic micro-motion.
Robust Outlier Suppression: A Hampel filter is applied to mitigate impulsive noise by replacing outliers with local medians.
Spectral Decoupling: The refined phase sequence undergoes FFT, followed by Gaussian smoothing and FIR band-pass filtering to separate respiration and heartbeat components. Heart rate is estimated via energy-weighted averaging within the cardiac band.

Key Contributions

Geometry-Regulated Embodied Sensing: The paper introduces a paradigm where radar observation geometry is actively regulated through feedback-driven pose alignment, enabling reliable sensing under unconstrained robot–human configurations.
Vision-Guided Pose Regulation with Robust Signal Extraction: A pipeline that steers the robot-mounted radar toward near-axial observation while integrating a phase-based enhancement module to stabilize extraction under pose variations.
Cross-Paradigm Validation: The system is evaluated against three sensing modes (Static, Proximity Embodied, and Geometry-Aware) and three baseline signal processing algorithms, demonstrating consistent accuracy improvements.

Experimental Results

The system was implemented on a Galaxea R1 Lite mobile robot equipped with a 60 GHz FMCW radar and an RGB-D camera. Experiments were conducted in a controlled indoor environment with seated participants.

Performance Metrics (Mean Absolute Error - MAE):

Breathing Interval (BI): ActiveVital reduced the error from 0.87 s (Proximity Embodied) to 0.14 s, achieving accuracy comparable to Static Sensing (0.15 s).
Heart Rate (HR): ActiveVital reduced the error from 13.59 bpm (Proximity Embodied) to 2.22 bpm. While Static Sensing achieved the lowest error (1.66 bpm), it required manual user alignment. ActiveVital achieved comparable performance without requiring subject cooperation.

Comparative Analysis:

Vs. Proximity Sensing: ActiveVital significantly outperformed proximity-based approaches, which suffered from high errors due to misalignment attenuating the cardiac signal.
Vs. Static Sensing: ActiveVital achieved similar accuracy to static sensing but with the critical advantage of autonomy, removing the need for users to maintain specific postures.
Algorithm Robustness: When tested against baseline algorithms (Second-Order DE, Band-pass Filter, Scattering Model), ActiveVital's geometry-aware approach maintained superior stability as geometric conditions deteriorated, showing the smallest performance degradation in Heart Rate estimation.

Significance and Claims

The paper claims that sensing pose is a controllable variable for robust non-contact physiological monitoring. By actively regulating the robot's pose to achieve near-axial alignment with the thoracic surface, ActiveVital maximizes the radial projection of cardiopulmonary micro-motion.

The authors assert that this approach:

Enables contactless monitoring in real domestic environments without wearables or pre-installed infrastructure.
Provides autonomous vital signs estimation that is robust to user posture variations, unlike static systems that rely on user cooperation.
Demonstrates that geometry-aware embodied sensing is essential for capturing weak cardiac signals, offering a solution where traditional proximity-based mobile sensing fails.

The work positions ActiveVital as an opportunistic function for multipurpose home robots, bridging the gap between task execution and continuous health monitoring.