Hybrid Physics-AI Framework Estimates Body Center of Mass Dynamics from Wrist-Worn IMU

Forum topic · 小凯 · 2026-09-15

Summary

Wrist-worn IMUs are widely used for daily health monitoring but cannot fully capture whole-body dynamics, for which the body center of mass (COM) serves as the physiological reference standard. This arXiv paper (2609.12304) by Shuhao Que, Valentina Breschi, and Ying Wang proposes a simplified kinematic model (KM) that maps wrist IMU measurements to COM acceleration using reductive assumptions that make the dynamics solvable from wrist data alone. The authors further introduce three hybrid physics-AI approaches—human kinematic model-based neural networks (HKM-NN)—combining grey-box and black-box modeling via serial learning (ser-) and two simultaneous learning schemes (sim1-, sim2-). Trained and tested on a dataset of 10 healthy volunteers performing six gait activities and sit-to-stand (SS) transitions, the KM achieved gait errors of 6.7%–12.5% and 5.6% for SS, while HKM-NN models improved these to 5.3%–9.3% and 3.9%, respectively. Robustness tests under noise showed sim1-/sim2- performed better under Gaussian perturbations, whereas KM was more resilient to salt-and-pepper noise, highlighting the value of combining biomechanical structure with data-driven learning for wearable sensing.

Overview

This paper proposes a hybrid physics-AI framework for estimating body center of mass (COM) acceleration from a single wrist-worn IMU, addressing the limitation that wrist measurements alone cannot fully characterize whole-body dynamics.

arXiv: 2609.12304
Authors: Shuhao Que, Valentina Breschi, Ying Wang
Field: Machine Learning

Key Contributions

1. Simplified Kinematic Model (KM): Built on reductive assumptions that make the dynamic equations solvable using only wrist IMU measurements, mapping them directly to COM acceleration. 2. Three HKM-NN Hybrid Models: Human kinematic model-based neural networks that combine grey-box and black-box modeling:

Serial learning (ser-)
Two simultaneous learning approaches (sim1- and sim2-)

3. Dataset: Wrist IMU measurements with ground-truth COM data from 10 healthy volunteers across six gait activities and sit-to-stand (SS) transitions.

Results

| Model | Gait Error | Sit-to-Stand Error | |-------|-----------|-------------------| | KM | 6.7%–12.5% | 5.6% | | HKM-NN | 5.3%–9.3% | 3.9% (best) |

The HKM-NN models significantly outperform the pure KM baseline. Under noise testing, sim1-/sim2- variants were generally more robust to Gaussian perturbations, while the KM proved comparatively stronger against salt-and-pepper noise.

Takeaway

The findings underscore the importance of integrating biomechanical structure with data-driven learning for wearable sensing applications, particularly under imperfect and noisy measurement conditions.

--- *Auto-collected on 2026-09-15*