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

Smart glasses can track where you are with only their built-in motion sensor, an inertial measurement unit (IMU); estimating position this way is called inertial odometry. It is low-power and privacy-friendly, but tiny errors accumulate: after minutes of walking, the estimate has drifted metres away.

MARIO starts from one observation: the IMU is attached to a person. A model infers from head motion how the body below is moving, keeping estimates consistent with human walking, and three sensors glasses already carry (a second IMU, a barometer, a magnetometer) steady them further.

The pose prior. A human-pose-grounded framework that injects a learnt IMU-inferred kinematics prior into the odometry model, improving stability and reducing drift.
The sensor fusion. A framework integrating barometer, magnetometer, and dual-IMU signals, substantially improving vertical and heading cues.
The evidence. Both modules drop into four existing architectures (TLIO, AirIO, RoNIN, EqNIO) with consistent gains across three datasets.

Why odometry drifts. A schematic, not measured data: each step adds a tiny heading error, and the errors compound along the walk. MARIO keeps the estimate tied to how a body actually moves, so the path stays close to the truth.

up to 44% less position error over long walks

up to 42% less drift as the walk gets longer

315 FPS inference speed, well beyond real time

Implementation

MARIO is not a new tracker built from scratch. It is two add-on modules that plug into existing inertial odometry systems; the same two modules improved all four systems we tried them on.

Architecture of the pose model. IMU acceleration and orientation streams enter a feature encoder followed by stacked GRU layers, which output joint positions and then joint rotations; a separate velocity decoder outputs joint velocities. — **The pose model.** A neural network reads the head’s acceleration and rotation and predicts how the **pelvis and legs** are moving. Its prediction anchors the tracker physically: position estimates must stay consistent with how a walking body actually moves. Arms and hands are left out; head motion says too little about them.

A block diagram of the sensor fusion pipeline: streams from the second IMU, barometer, and magnetometer are time-aligned with the primary IMU, each encoded into a compact feature, then concatenated with the pose model's prediction and fed into the host inertial odometry model. — **Sensor fusion.** Each additional sensor stream is aligned in time with the main IMU and compressed into a compact summary of what it is currently reporting. The summaries are then fed into the host tracker together with the pose model’s prediction, so any of the four trackers can use them without changing its own design.

The pose model. A neural network reads the head’s acceleration and rotation and predicts how the pelvis and legs are moving. Its prediction anchors the tracker physically: position estimates must stay consistent with how a walking body actually moves. Arms and hands are left out; head motion says too little about them.

Head motion in, pelvis and legs out.

Inside the pose model: a shared GRU feeds a position branch, a rotation branch and a velocity decoder.

Sensor fusion. Each additional sensor stream is aligned in time with the main IMU and compressed into a compact summary of what it is currently reporting. The summaries are then fed into the host tracker together with the pose model’s prediction, so any of the four trackers can use them without changing its own design.

Every stream is encoded, concatenated, fused, then handed to the existing tracker.

Each fused sensor fixes a different kind of error:

Barometer: senses air pressure, which falls as you climb. This anchors altitude on stairs and slopes.
Magnetometer: senses the Earth’s magnetic field, giving a compass heading that stops the estimated path from slowly rotating away from the real one.
Second IMU: sits a few centimetres from the first, so any head rotation reads slightly differently on each. Comparing the two pins rotation down more firmly.

Barometer. Air pressure falls as you climb, which anchors altitude on stairs and slopes.

Magnetometer. The Earth’s field gives a compass heading, so the path stops rotating away.

Second IMU. A few centimetres from the first, it reads head rotation slightly differently.

How we tested it

We trained and evaluated on Nymeria: 264 people wore sensor-equipped glasses for about 300 hours and 400 kilometres across 50 environments, with motion capture recording their true position.

264people wearing sensor-equipped glasses

~300 hof recorded motion

400 kmwalked in total

50different environments

Evaluation walks are held out from training, and we also ran the model, without any retraining, on a second dataset of home activities. We report how far the estimate ends up from the truth and how fast that error grows.

Results

The pose model alone already reduces error. Adding the fused sensors helps most on long walks, exactly where inertial odometry drifts the worst.

10.19 m → 5.73 m average position error of TLIO on Nymeria, before and with MARIO

4 / 4 trackers improved by the same two add-on modules

up to 53% less drift on home recordings it never trained on

All four trackers improved:

TLIO
AirIO
RoNIN
EqNIO

Average position error on Nymeria. The same two add-on modules, on TLIO and AirIO, two of the four trackers.

AirIO’s error fell from 6.85 metres to 4.64, and the remaining two, RoNIN and EqNIO, improved similarly, so the idea is not tied to any one design. On unseen home recordings, error still dropped by roughly a third, and inference runs comfortably in real time.

Removing one sensor at a time shows they are not redundant: the barometer contributes most to vertical accuracy, while the magnetometer and second IMU remove much of the remaining drift.

Team

This is research from the SPICE Lab at Northwestern University by:

Publication

This work appears at CVPR 2026 (Findings). The published version is available via its DOI, and the authors’ print is available here (PDF).

Li, Y., Yeon, T., Gao, C., Xu, V., Liu, X., & Ahuja, K. (2026). MARIO: Motion-Augmented Real-Time Multi-Sensor Inertial Odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (Findings), pp. 3460–3469. https://doi.org/10.1109/CVPRF71909.2026.00336

@inproceedings{li2026mario,
  title={MARIO: Motion-Augmented Real-Time Multi-Sensor Inertial Odometry},
  author={Li, Yiquan and Yeon, Taeyoung and Gao, Chenfeng and Xu, Vasco and Liu, Xuanyou and Ahuja, Karan},
  booktitle={2026 IEEE/CVF Conference on Computer Vision and Pattern Recognition - Findings Track (CVPRF)},
  address={Denver, CO, USA},
  pages={3460--3469},
  year={2026},
  doi={10.1109/CVPRF71909.2026.00336}
}

MARIO

Implementation

How we tested it

Results

Team

Publication

MARIO

Walk long enough with only an IMU, and the estimate drifts metres away.

The IMU is attached to a person, so use what a body can do.

Infer the body from the head

Add what glasses already carry

Head motion goes in, the motion of pelvis and legs comes out.

Every sensor stream is encoded, then fused with the pose prediction.

Each fused sensor fixes a different kind of error.

Trained and tested on Nymeria, then run on home recordings with no retraining.

Up to 44% less position error, and drift grows more slowly.

The same two modules improved all four trackers.

Know the body, and the tracker drifts less.

Removing one sensor at a time shows they are not redundant.

Vertical accuracy

Heading

Rotation

The numbers behind the headline, from the paper.

Overview

Implementation

How we tested it

Results

Team

Publication