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

Forum topic · 小凯 · 2026-09-15

Summary

A paper by Shuhao Que, Valentina Breschi, and Ying Wang (arXiv:2609.12304) proposes a hybrid physics-AI framework that estimates whole-body center of mass (COM) acceleration from a single wrist-worn IMU. The authors first derive a simplified kinematic model (KM) whose reductive assumptions make the dynamics solvable from wrist measurements alone, then introduce three hybrid HKM-NN methods combining grey-box and black-box modeling: serial learning (ser-) and two simultaneous learning variants (sim1- and sim2-). Models were trained and tested on data from 10 healthy volunteers performing six gait activities and sit-to-stand transitions, with ground-truth COM measurements. The KM alone achieved 6.7%-12.5% error on gait and 5.6% on sit-to-stand; HKM-NN models improved this to 5.3%-9.3% on gait and 3.9% on sit-to-stand. Under noise, sim1-/sim2- were more robust to Gaussian perturbations, while the KM handled salt-and-pepper noise better, showing the value of blending biomechanical structure with data-driven learning for wearable sensing.

Overview

This post summarizes an ML paper by Shuhao Que, Valentina Breschi, and Ying Wang, published on arXiv (2609.12304).

Wrist-worn IMUs are ubiquitous in daily-life health monitoring, but they cannot fully characterize whole-body dynamics—the body center of mass (COM) is the recognized physiological reference standard. The authors therefore propose a simplified kinematic model (KM) that maps wrist IMU measurements to COM acceleration, built on reductive assumptions that make the dynamic equations solvable from wrist data alone.

Key contributions

A simplified kinematic model (KM) for estimating COM acceleration from a wrist-worn IMU.
Three hybrid AI modeling approaches (HKM-NN) combining grey-box and black-box modeling:
Serial learning (ser-)
Simultaneous learning (sim1-)
Simultaneous learning (sim2-)

Dataset and results

Models were trained and tested on a dataset collected by the authors: 10 healthy volunteers performing six gait activities and sit-to-stand (SS) transitions, with wrist IMU measurements and ground-truth COM measurements.

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

Robustness findings

Under noisy test conditions, the hybrid variants showed different characteristics: sim1-/sim2- HKM-NN models were generally more robust to Gaussian perturbations, while the KM model performed relatively better under salt-and-pepper noise. The results highlight the importance of combining biomechanical structure with data-driven learning for wearable sensing under imperfect, noisy measurement conditions.

---

*Paper link: arXiv:2609.12304*