Title: MARIO: Motion-Augmented Real-Time Multi-Sensor Inertial Odometry

Abstract

For augmented reality (AR) applications and wearable technology, Inertial Odometry (IO) reliant solely on Inertial Measurement Units (IMUs) offers a streamlined approach to tracking human movement. While recent learning-driven IO techniques have enhanced the generalizability of inertial localization by leveraging large-scale pretraining on human motion datasets, these models still struggle with noise and drift. This vulnerability persists because they often fail to explicitly account for human motion dynamics, particularly when applied to daily activity datasets like Nymeria.

To address these limitations, we propose anchoring inertial odometry in human kinematics via a learned IMU-inferred pose prior. This mechanism enforces physically consistent motion constraints. By embedding this pose prior into current IO architectures, we achieved a positional drift reduction of up to 36% on the Nymeria dataset—a benchmark five times larger than those utilized in previous studies.

Furthermore, we enhanced long-term stability through a sensor-fusion framework that integrates auxiliary data from lightweight sensors commonly found in commercial AR glasses, such as magnetometers, barometers, and secondary IMUs. This multimodal fusion strategy lowered positional drift by as much as 42%, significantly boosting robustness and generalization across varied motion scenarios. Our findings establish a new standard for accurate, camera-less human tracking by merging human motion kinematics with multimodal sensing, thereby defining a fresh paradigm for inertial and lightweight odometry.

For more details, visit our project page at https://spice-lab.org/projects/MARIO/.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC