Learned Inertial Odometry for Cycling Based on Mixture of Experts Algorithm

Hao Qiao, Yan Wang, Shuo Yang, Xiaoyao Yu, Jian Kuang, Xiaoji NiuarXiv 2025 · DOI · arXiv

Introduction

Reliable positioning is a cornerstone of modern urban mobility, particularly for the rapidly expanding bicycle-sharing industry. While Global Navigation Satellite Systems (GNSS) like GPS are ubiquitous, their performance degrades significantly in "urban canyons," tunnels, or under dense foliage. For bicycles, maintaining accurate localization during these outages is essential for fleet management, user navigation, and safety.

Traditional solutions rely on Inertial Navigation Systems (INS) using accelerometers and gyroscopes. However, these systems are prone to cumulative "drift" errors. While recent deep-learning approaches have successfully mitigated drift by predicting velocity directly from sensor data, they often demand computational resources far exceeding the capabilities of low-power mobile devices or e-bike hardware. This paper introduces a novel, lightweight navigation system that leverages a Mixture of Experts (MoE) architecture to achieve high-precision 3D positioning with a fraction of the computational cost.

The Challenge of Edge-Based Inertial Odometry

The primary hurdle in deploying AI-driven inertial odometry on bicycles is the "compute budget." State-of-the-art models like TLIO (Tight-Coupled Learned Inertial Odometry) provide impressive accuracy but are computationally "heavy." Bicycles and e-bikes typically utilize low-power processors where battery life and real-time latency are critical.

Furthermore, bicycles present unique physical challenges. Many traditional models rely on "non-holonomic constraints"—mathematical assumptions that the vehicle does not slide or move sideways. While these hold true for cars on dry pavement, they fail for bicycles on unpaved trails, gravel, or slippery surfaces. A robust system must learn the complex dynamics of cycling across diverse terrains without being tethered to rigid mathematical assumptions.

Methodology: Efficiency through Specialization

The proposed framework combines the physical rigor of classical robotics with the adaptive power of modern deep learning.

1. Hybrid Backbone: EKF and Neural Networks

The system utilizes an Error-State Extended Kalman Filter (EKF) as its backbone. The EKF manages the fundamental physics of motion, while a neural network provides "corrections" by predicting the bicycle's instantaneous velocity and the associated uncertainty (covariance). This hybrid approach ensures that the model remains grounded in physical reality while benefiting from data-driven insights.

2. Mixture of Experts (MoE) Architecture

The core innovation lies in the use of a Mixture of Experts (MoE) architecture. Rather than employing a single, monolithic neural network, the system uses a "team" of smaller, specialized expert networks.

  • Gating Network: A manager network analyzes incoming sensor data to determine which experts are best suited for the current motion—for instance, distinguishing between a sharp turn and steady cruising on a flat road.
  • Shared-Expert Strategy: Inspired by recent advancements in Large Language Models, the architecture includes a "shared expert" that remains active to capture general motion patterns.
  • Top-K Activation: Only a small subset (Top-K) of specialized experts is activated at any given time. This "sparse activation" allows the model to maintain a high total parameter count for learning complex behaviors while only "firing" a small percentage of neurons during inference.
Experimental setup showing a smartphone mounted on a bicycle handlebar for data collection.

Data Collection and Diversity

To train and validate the MoE model, the researchers curated a comprehensive dataset specifically for cycling:

  • Scale and Scope: The dataset comprises over 25 hours of sensor data collected by 8 different participants.
  • Environmental Variety: Testing occurred across various bicycle models and road conditions, ranging from smooth paved city streets to bumpy, unpaved trails.
  • Ground Truth: High-accuracy reference trajectories were established using a Velocity Dead Reckoning (VDR) algorithm, providing a reliable baseline for the AI to learn from.

Performance and Results

The MoE-based model was benchmarked against LLIO, a leading lightweight baseline. The results demonstrate a significant leap in both efficiency and robustness.

Computational Gains

The MoE architecture achieved a staggering 81.8% reduction in FLOPs (floating-point operations) compared to the baseline. Additionally, the model size was reduced by 64.7%, bringing the total parameter count down to just 1.2 million. This makes the system exceptionally well-suited for deployment on edge devices.

Accuracy and Robustness

Despite the drastic reduction in computation, the model outperformed the baseline in terms of inference error (0.205 vs. 0.272). Most notably, the model demonstrated superior performance on unpaved surfaces. While traditional physics-based models often fail when mathematical constraints are violated by bumps or slips, the learned MoE model successfully adapted to these "messy" real-world conditions.

Comparison of predicted trajectories (blue) vs. ground truth (red) across paved and unpaved conditions.

Conclusion and Future Directions

This research proves that the Mixture of Experts architecture is not just for massive language models; it is a powerful tool for downscaling AI for specialized sensor fusion tasks. By activating only the necessary "neurons" for specific cycling maneuvers, the system provides high-accuracy navigation that is both robust and incredibly efficient.

Future work will focus on further model compression through knowledge distillation and direct hardware validation to measure real-time battery consumption on commercial e-bike controllers. As urban mobility continues to evolve, such efficient, AI-driven navigation will be vital for the next generation of smart cycling infrastructure.

Summary updated

Abstract

GNSS suffers from multipath errors in urban canyons, making reliable bicycle localization difficult. Hand-crafted inertial alternatives, such as cycling dead reckoning and nonholonomic constraints, fail to generalize across riders, postures, and road surfaces. This work implements learned inertial odometry for bicycle localization: a neural velocity predictor, trained on multirider, multi-bicycle, and multi-surface IMU data, is tightly coupled with an extended Kalman filter for three-dimensional pose estimation, replacing hand-crafted cycling models with a data-driven representation. To reduce computational overhead, the predictor is realized as a sparsely-gated Mixture of Experts (MoE) network with top-K routing, trained under an alternating parameter-freezing scheme with a per-expert capacity constraint to encourage diverse and balanced specialization.The proposed model attains 0.333 m/s inference error, 9.49 m ATE, and 2.58 m RTE at only 28.76 M FLOPs. This corresponds to approximately 7 times and 9 times fewer FLOPs than the ResMLP and ResNet backbones used in LLIO and TLIO, while maintaining comparable or superior accuracy. The proposed MoE variants further rank first on both ATE and RTE across all evaluation splits, demonstrating strong generalization.
BibTeX
@misc{qiao2025learned,
  author = {Hao Qiao and Yan Wang and Shuo Yang and Xiaoyao Yu and Jian Kuang and Xiaoji Niu},
  title = {{Learned Inertial Odometry for Cycling Based on Mixture of Experts Algorithm}},
  year = {2025},
  eprint = {2510.17604},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.17604},
}
Figure