---
title: 'ActiveVital: Geometry-Aware Vital Monitoring'
url: https://www.emergentmind.com/topics/vital
type: topic
---

# ActiveVital: Geometry-Aware Vital Monitoring

ActiveVital is a geometry-aware embodied vital signs monitoring framework for home healthcare robots that reformulates non-contact respiration and heart-rate sensing from passive signal recovery into active regulation of sensing geometry. The system is built around a robot-mounted $60\ \text{GHz}$ FMCW mmWave radar and an RGB-D camera, and treats the radar–chest incidence angle as an explicit control variable. Its central claim is that, in unconstrained home settings, vital-sign observability is dominated not only by signal processing but by robot–human geometry, because mmWave radar observes only the radial component of thoracic motion. ActiveVital therefore localizes a visually derived chest anchor, drives the radar toward near-normal incidence on the thoracic surface, and then refines the extracted slow-time phase with temporal differencing and Hampel filtering to stabilize respiration and heartbeat estimation under embodied motion [2606.30275].

## 1. Problem setting and conceptual shift

Non-contact vital signs monitoring in homes is difficult because users may be seated, sleeping, or moving with arbitrary orientations, furniture may occlude parts of the body, and the robot’s relative pose varies over time. In these conditions, mmWave radar remains attractive because its phase is sensitive to sub-millimeter motion, but the measurement is fundamentally limited to the line-of-sight component of displacement. Angular misalignment between the radar boresight and the thoracic surface normal therefore attenuates the measurable phase from respiration and, more severely, from heartbeat, whose amplitude is smaller [2606.30275].

ActiveVital addresses this by shifting the problem formulation. Rather than assuming that better DSP alone can recover weak physiological signals from arbitrary poses, it treats sensing geometry as a controllable state in a perception–action loop. This produces a chest-centric alignment behavior in which the robot actively steers the radar toward near-normal incidence, thereby maximizing radial observability. A plausible implication is that ActiveVital belongs to a broader class of embodied sensing systems in which perception quality is regulated through motion rather than post hoc compensation.

This framing is consistent with broader mmWave vital-sign literature showing strong sensitivity to distance, angle, and orientation. A public mmWave FMCW radar dataset for adult vital-sign monitoring reports that performance is best at $0^\circ/80\ \text{cm}$ and degrades with angle and non-frontal orientation, with distance and angle reducing SNR according to the antenna gain pattern [2405.12659]. ActiveVital’s contribution is to close the loop on that geometry rather than treating it as a nuisance parameter.

## 2. Geometry and radar measurement model

The geometric core of ActiveVital is the phase–range relationship of mmWave radar. Let the wavelength be $\lambda$ at carrier frequency $f_0$, and let the line-of-sight range be
$$
R(t)=R_0+\Delta R(t).
$$
The measured phase is
$$
\phi(t)= \frac{4\pi}{\lambda} R(t) + \phi_0,
$$
so phase changes follow
$$
\Delta \phi(t)= \frac{4\pi}{\lambda} \Delta R(t).
$$
If thoracic displacement is represented by a vector normal to the chest surface, with unit surface normal $\mathbf{n}$ and radar boresight unit vector $\mathbf{b}$, then the incidence angle is
$$
\theta = \arccos(\mathbf{n}\cdot \mathbf{b}),
$$
and the radial projection is
$$
m_r(t)= |m(t)| \cos \theta,
$$
with $\Delta R(t)=m_r(t)$. Observability and SNR are therefore maximized when $\theta \approx 0$, i.e., near-normal incidence [2606.30275].

The underlying radar is FMCW. ActiveVital specifies transmit and received signal models in which the chest echo delay is $\tau(t)=2R(t)/c$, and the baseband IF phase tracks chest motion after mixing and range FFT. With multiple scattering centers under a GTD approximation, the IF phase term for the $i$-th scatterer is
$$
\phi_{if}^{(i)}(t) \approx 2\pi f_0 \tau_i + 2\pi K_s \tau_i t - 2\pi f_d^{(i)} t.
$$
After selecting the chest range bin $j_{trgt}$, the slow-time phase is
$$
\phi(m)= \arg(S_{IF}(j_{trgt}, m)), \quad m=1\dots M.
$$

The modeling assumptions are explicit: small-motion linearization, far-field or paraxial approximation so the chest patch is locally planar, and the presence of multipath but with chest-centric alignment emphasizing dominant specular or near-specular scatterers. This is closely aligned with mmWave FMCW formulations used in radar vital-sign datasets, which also use
$$
\varphi(t)=\frac{4\pi}{\lambda} d(t)
$$
to map phase to displacement and identify subject range bins via range FFT [2405.12659].

## 3. Vision-guided chest localization and active geometric regulation

ActiveVital estimates a 2D chest anchor from human pose keypoints detected by ViTPose. Using left and right shoulders $P_{ls}, P_{rs}$ and left and right hips $P_{lh}, P_{rh}$, it computes shoulder and hip midpoints
$$
P_{sc}=\frac{P_{ls}+P_{rs}}{2}, \qquad P_{hc}=\frac{P_{lh}+P_{rh}}{2}.
$$
The vertical chest anchor is interpolated along the torso axis as
$$
P_{y,ref}= \alpha P_{y,sc} + (1-\alpha) P_{y,hc},
$$
with $\alpha=0.75$, producing an upper-chest bias. The shoulder width is
$$
W_{shoulder}= |P_{x,ls}-P_{x,rs}|,
$$
and the horizontal coordinate is offset toward the cardiac side:
$$
P_{x,ref}= P_{x,sc} + \beta W_{shoulder}, \qquad \beta=0.1.
$$
The final anchor is
$$
P_{ref}= [P_{x,ref}, P_{y,ref}]^T
$$
[2606.30275].

Depth is estimated around the anchor as the median value in a neighborhood:
$$
D_{target}= \operatorname{median}\{D(u,v)\mid (u,v)\in N(P_{ref})\}.
$$
With camera intrinsics $(f_x,f_y)$ and image center $(L_x,L_y)$, pixel deviations are
$$
e_x = x-L_x,\qquad e_y = y-L_y,
$$
and these are mapped to metric displacements:
$$
\Delta H= a_v |e_y| D_{target}/f_y,\qquad
\Delta L= a_h |e_x| D_{target}/f_x,
$$
where $a_h,a_v\in\{-1,0,1\}$ are motion directions determined by deadband logic. The control loop repeatedly detects $P_{ref}$, computes pixel error $E=[e_x,e_y]$, maps that error to metric commands, and actuates the arm until the radar boresight intersects the chest anchor, driving $\theta \to 0$ [2606.30275].

The implemented controller uses binary sign control with deadbands:
$$
a_v=
\begin{cases}
1 & e_y>\delta_y\\
-1 & e_y<-\delta_y\\
0 & \text{otherwise}
\end{cases}
$$
with an analogous definition for $a_h$, while depth is regulated by
$$
\Delta D = D_{target}-D_{ref}.
$$
The paper notes a representative PID option if explicit angular error can be estimated, but this is not required by the implementation. Alignment is considered complete when $|e_x|$, $|e_y|$, and $|\Delta D|$ all fall within thresholds for a short hold window, after which radar data are collected.

## 4. Differential phase enhancement and vital-sign estimation

After geometric alignment, ActiveVital processes the radar slow-time phase. It unwraps the extracted phase to obtain $\phi_{unwrap}(m)$ and then applies first-order temporal differencing:
$$
\Delta \phi_{unwrap}(m)= \phi_{unwrap}(m)-\phi_{unwrap}(m-1).
$$
This suppresses drift, attenuates slow baseline wander, and sharpens periodic respiration and heartbeat oscillations [2606.30275].

To suppress impulsive noise, ActiveVital applies a Hampel filter over a window $\Omega_m=\{m-r,\dots,m+r\}$. It computes
$$
med(m)= \operatorname{median}\{\Delta \phi_{unwrap}(i)\},
$$
$$
MAD(m)= \operatorname{median}|\Delta \phi_{unwrap}(i)-med(m)|,
$$
and
$$
\sigma(m)= \kappa MAD(m), \qquad \kappa=1.4826.
$$
Samples violating
$$
|\Delta \phi_{unwrap}(m)-med(m)|>\tau \sigma(m)
$$
are replaced by the median, yielding a refined sequence $\tilde{\phi}(m)$. The authors emphasize that this refinement is temporal rather than inter-channel or inter-beam differential enhancement, and report that it proved more robust under embodied motion [2606.30275].

The downstream pipeline is conventional but tightly integrated with the geometric loop. Range FFT produces the range profile, a chest bin $j_{trgt}$ is selected via amplitude and temporal continuity, and the refined slow-time phase is transformed spectrally:
$$
P(f)= |\operatorname{FFT}\{\tilde{\phi}(m)\}|.
$$
Gaussian smoothing and FIR band-pass filters isolate respiration, typically $0.1$–$0.5\ \text{Hz}$, and heart activity, typically $0.8$–$3\ \text{Hz}$. Heart rate is estimated through an energy-weighted peak frequency in the cardiac band:
$$
f_{heart}= \frac{\sum_k f_k P_{heart}(f_k)}{\sum_k P_{heart}(f_k)}, \qquad
HR = 60 f_{heart}.
$$
Respiration is estimated in the time domain from detected peaks:
$$
BR = \left(\frac{K-1}{t_K^{resp}-t_1^{resp}}\right)\times 60,
$$
with breath interval defined as the mean peak-to-peak time. A sliding-window heart-rate estimator uses $20\ \text{s}$ initialization and $2\ \text{s}$ updates [2606.30275].

## 5. System architecture, implementation, and experimental results

The implemented system mounts a $60\ \text{GHz}$ FMCW radar and an RGB-D camera on the arm of a Galaxea R1 Lite mobile platform. The radar is specified as $57.5$–$63.5\ \text{GHz}$, $1\text{T}3\text{R}$, chirp slope $91.5\ \text{MHz}/\mu s$, ADC $2\ \text{MHz}$, chirp duration $71.8\ \mu s$, frame period $50\ \text{ms}$, and range resolution approximately $0.0228\ \text{m}$. Camera inference drives alignment, after which radar capture proceeds at approximately $20\ \text{Hz}$ equivalent with short holds for stable measurement [2606.30275].

Experiments were conducted in a $6.6\ \text{m} \times 6.3\ \text{m} \times 3.5\ \text{m}$ indoor space. Subjects were seated naturally without alignment instructions. The robot started at approximately $1\ \text{m}$ from the subject and aligned within a few seconds. Ground truth was provided by a fingertip pulse oximeter for heart rate and an external RGB camera tracking chest motion for respiration [2606.30275].

Three paradigms were compared: Static Sensing, in which the radar was fixed and the subject manually aligned; Proximity Embodied Sensing, in which the robot approached but did not align chest geometry; and Geometry-Aware Embodied Sensing, i.e., ActiveVital. Using the same signal processing across paradigms, cross-paradigm error statistics were:

| Paradigm | BI MAE/RMSE | HR MAE/RMSE |
|---|---:|---:|
| Static | 0.15 / 0.19 s | 1.66 / 1.96 bpm |
| Proximity | 0.58 / 0.78 s | 5.26 / 6.59 bpm |
| ActiveVital | 0.14 / 0.17 s | 2.22 / 2.59 bpm |

The headline comparison reports respiration interval error reduced from $0.87\ \text{s}$ to $0.14\ \text{s}$ and heart-rate error reduced from $13.59\ \text{bpm}$ to $2.22\ \text{bpm}$ when moving from proximity-like passive filtering to ActiveVital [2606.30275].

Over $30$ independent $25\ \text{s}$ recordings, ActiveVital achieved respiration-rate MAE $0.26\ \text{bpm}$, RMSE $0.33\ \text{bpm}$, and P95 $0.72\ \text{bpm}$; heart-rate MAE $1.67\ \text{bpm}$, RMSE $2.16\ \text{bpm}$, and P95 $4.65\ \text{bpm}$. Spectral analyses show that off-axis and long-range configurations produce indistinct peaks, while chest-centered alignment yields a clear respiration fundamental, harmonic structure, and heartbeat peaks, supporting the claim that driving $\theta \to 0$ increases SNR for cardiac micro-motion [2606.30275].

## 6. Relation to adjacent work, limitations, and significance

ActiveVital sits at the intersection of contactless radar monitoring, embodied robotics, and geometry-aware control. Related work establishes each ingredient separately. A public mmWave FMCW radar dataset demonstrates adult vital-sign monitoring across distance, angle, orientation, and extreme physiological scenarios, and provides standardized raw radar, HR, and RR references for benchmarking signal-processing pipelines [2405.12659]. MERIT shows that multimodal radar–IMU fusion can reconstruct ECG waveforms during daily motion, emphasizing that motion robustness often requires multimodality rather than a single sensing stream [2410.00392]. VitalLens, in contrast, uses selfie video for real-time heart-rate and respiration-rate estimation, showing that camera-based methods can also reach low error under favorable conditions but remain sensitive to movement and illuminance variation [2312.06892].

ActiveVital’s limitation is not hidden by its results. The radar remains a radial-only sensor, so non-radial chest motion is not observed; strong angular offsets or occlusions still degrade the signal. Heavy occlusion, unfavorable clothing or blanket layering, and multiperson scenes complicate target selection. The small-motion assumption can also fail under extreme macro-motion, breaking phase-unwrapping continuity [2606.30275]. These are structural rather than incidental limitations.

The paper identifies several future directions: multi-view fusion with multiple radars or cameras for robust thoracic normal estimation and direct $\theta$ control, learning-based denoising combined with GTD chest-scattering priors, predictive control such as MPC to maintain alignment while the user moves, and multimodal integration with UWB radar, camera-based PPG, or thermal sensing [2606.30275]. This suggests a broader research program in which physiological observability becomes an actively maintained property of robot behavior rather than a passive assumption.

In that sense, ActiveVital’s distinctive contribution is not merely improved estimation error, but a redefinition of non-contact robotic vital-sign monitoring as a closed-loop geometric regulation problem. The system approaches static-sensing accuracy while operating under unconstrained robot–human configurations, and does so by making alignment itself part of the sensing method rather than a preprocessing convenience [2606.30275].

Source: https://www.emergentmind.com/topics/vital