Source-linked AI summary
One Residual with Three Reuses: A Wristband Front End for Gesture Sensing
Sam Rifaki
TL;DR
Coin-cell wristband sensing must coordinate gesture detection, sensor routing, and drift-robust tracking while combining IMU and 60 GHz FMCW radar. This design study uses one on-chip residual generator for all three functions, achieving 47% lower classifier invocation energy at 90% recall and 4.6× lower pose-tracking error under measurement bias drift.
Problem
Coin-cell wristbands need coordinated classifier gating, modality routing, and drift compensation, but separate calibration tables and silicon area limit shared-MCU architectures.
Method
A single on-chip residual generator uses prediction error from conditioned IMU and radar streams to gate classification, select modalities, and reweight EKF measurements.
Results
47% lower classifier invocation energy at 90% recall and 4.6× lower pose-tracking RMS error under bias drift were achieved, while detection probability reached 0.72/0.80 at 1% false-alarm rate.
Takeaways & Limitations
The design study demonstrates that one residual generator can support the wristband’s gating, routing, and drift-aware tracking functions within a compact MCU implementation.
Takeaways & Limitations
Measured silicon power and on-body capture are deferred to follow-on hardware studies.
Abstract
from arXiv · showhide
Continuous wrist-worn hand sensing for gesture interfaces and motor symptom monitoring needs an always-on front end that fits inside a coin-cell power budget while pairing a micro-electro-mechanical-systems (MEMS) inertial measurement unit (IMU) with a 60 GHz frequency-modulated continuous-wave (FMCW) radar to stay robust under occlusion and on-body drift. We present a design study of such a wristband front end in which classifier wake-up gating, mmWave versus IMU routing, and innovation-based EKF measurement reweighting share a single on-chip residual generator. The shared generator occupies 14.4 KB of program memory and 278 B of state and runs at 110K multiply-accumulates (MACs) per frame on an Ambiq Apollo4 Blue Plus class edge microcontroller unit (MCU). Across four public sensor data corpora (IPN Hand, SHREC 2021, MiliPoint 60 GHz FMCW radar, EAT-Radar) the front end reaches detection probability $P_D = 0.72/0.80$ at a 1% false-alarm rate, sustains a 47% classifier invocation energy reduction at 90% gesture detection recall, and lowers pose tracking root-mean-square error by $4.6\times$ under measurement bias drift relative to an adaptive Kalman with $R$-inflation baseline. Measured silicon power and on-body capture are deferred to follow-on hardware; the contribution here is a design study.
I. INTRODUCTION
Wrist-worn hand sensing combines MEMS IMUs with 60 GHz FMCW radar to remain robust under occlusion and on-body variation, but separate function-specific subsystems exceed coin-cell wristband constraints. This work unifies classifier wake-up, modality selection, and EKF measurement reweighting through one on-chip residual generator.
- Motivation: MEMS IMU–60 GHz FMCW radar combinations recover hand pose under occlusion, varying on-body placement, and lighting changes that defeat optical-only systems,.These systems support gesture interfaces, rehabilitation tracking, and clinical motor-symptom assessment.
- Problem: Separate wake-up, drift-tolerance, and modality-switching subsystems require independent calibration tables and silicon area, limiting suitability for coin-cell wristbands sharing one MCU and residual path.Prior approaches include online change-point testing, Kalman-style innovation residuals, interacting multiple-model or hybrid learning-based filters,, and information-gain scheduling,.
- Contribution: A single on-chip residual generator combines classifier wake-up gating, next-frame modality selection, and EKF measurement reweighting.It uses the one-step prediction error ξt from a compact recurrent forecaster operating on conditioned IMU and radar streams.
II. SYSTEM DESIGN AND IMPLEMENTATION
The wristband combines a 60 GHz FMCW radar, MEMS IMU, and MCU-resident recurrent forecaster into one residual-driven front end. The residual gates classification, routes modalities, and re-estimates EKF measurement noise within a constrained MCU budget.
- Signal chain: A single recurrent forecaster on the MCU predicts the next conditioned frame vector from conditioned IMU and mmWave streams, producing the residual used by downstream control blocks.The radar forwards on-die range-FFT bins at 30 frames/s, while the IMU streams raw accelerometer and gyroscope frames at 100 Hz over I2C.
- Three reuses of one signal: The shared residual supports three reuses: wake-up gating, per-modality routing, and EKF measurement-noise re-estimation.The wake-up threshold is fixed-point and calibrated for a false-alarm rate below 1%; the EKF uses an EWMA of residual energy to scale each modality’s R_k.
- Three reuses of one signal: The router selects the modality with minimum per-modality residual, matching the innovation-covariance ranking used by classical sensor schedulers –.Under linear-Gaussian observations, this choice minimizes expected squared residual plus measurement noise covariance.
- MCU budget: Quantization at h=32 yields 14.4 KB of program memory, 278 B of state, and 110 K MACs per frame for the deployment generator.The target is an Ambiq Apollo4 Blue Plus Cortex-M4F at 96 MHz; public specifications report about 4 µA/MHz active-mode draw at nominal voltage and operation down to roughly 0.5 V [11].
III. RESULTS
The shared-residual front end achieves strong wake-up and drift-tracking results while reducing routing energy and fitting within a small embedded memory budget. Its tracking block reduces RMS error against two filtering baselines, and the architecture uses substantially fewer MACs than recent embedded gesture pipelines.
- Wake-up gating: P_D = 0.72 at P_FA = 0.01 on IPN Hand and P_D = 0.80 on SHREC 2021 at the 90% recall operating point.The results exceed innovation-residual baseline by +0.09 and +0.12, and online change-point baseline by +0.17 and +0.21, respectively.
- Modality routing: 69.4% lower per-frame energy at P_D = 0.722 ± 0.005, using 30.6% of total ensemble MACs on 30,000 IPN Hand frames.This routing performance is within ±0.005 of a supervised oracle gate; a reduced-complexity gate cuts routing latency by 2.47× at a 0.003 P_D cost.
- Drift-aware tracking: 4.6× lower RMS tracking error at b_s = 5 cm versus adaptive Kalman noise re-estimation, and 2× lower versus a hybrid IMM-style filter.These reductions come from the EKF measurement noise re-estimation block.
- Comparison: Roughly 4–50× fewer MACs than several recent measured embedded gesture pipelines,, while fitting within 14.4 KB of program memory on one Cortex-M4F-class part.Table I compares MCU MAC counts and memory requirements across wearable mmWave-and-IMU front ends; this work is a design study combining wake-up, routing, and drift tracking with one residual.
IV. FAILURE MODES AND OPERATING ENVELOPE
The shared-residual architecture has three operational envelopes: post-selection adversarial corruption, bias drift faster than EWMA tracking, and heavy-tailed measurement noise. These conditions can suppress routing or diagnostic signals and force fallback behavior, while proposed remedies remain future work.
- Adversarial corruption: Noise injected after modality selection can defeat reactive selection when its energy exceeds five times the nominal noise floor.The attack raises ξ_t on the corrupted channel only after selection.
- Bias drift: Bias oscillations faster than the EWMA window reduce diagnostic signals, degrading RMS performance toward the camera-only reference below five EWMA windows.Tracking follows bias frequencies of order 1/(Nα); calibration values N and α are stored in flash and cannot change dynamically during deployment.
- Heavy-tailed noise: Student-t measurement noise with degrees of freedom near 3 increases R_k globally, suppressing routing signals and forcing the camera-only fallback.A Huber-clipped innovation norm is identified as a remedy but reserved for future work.
- Scope: The study makes no extrapolation beyond observed behavior for the stated operational envelopes.
V. CONCLUSION
The study presents a coin-cell-powered wristband front end combining mmWave and IMU sensing with three residual-generator-driven functions, achieving detection and energy-saving targets while deferring silicon-power measurement.
- V. CONCLUSION: PD = 0.72/0.80 at PFA = 0.01, with a 47 % reduction in classifier invocation energy per frame at 90 % recall.The full system integrates wake-up gates, modality routers, and EKF measurement reweighting generated by a single 278 B on-chip residual generator.
- V. CONCLUSION: The design study provides a coin-cell-powered wristband front end combining mmWave and IMU sensing with wake-up gating, modality routing, and EKF measurement reweighting.Future work includes measuring silicon power on an Apollo-class MCU.