🛰️ When GNSS Fails, WheelRoute Keeps Robots on Track

WheelRoute: Route-Constrained Robust Localization for Wheeled Robots in GNSS-Degraded Environments

Official paper: Route-Constrained Robust Fusion Estimation for MEMS/GNSS Integrated Navigation of Unmanned Ground Vehicles in GNSS-Degraded Environments

Jingzhi Cui¹ · Chao Zhang² · Yuliang Mao² · Shaolin Lü¹ · Dongmei Li¹ · Huan Che² · Rong Zhang¹

¹Tsinghua University    ²Xiaomi Inc.

GNSS drops in the tunnel. Dead reckoning drifts away. WheelRoute pulls the robot back to the route.

---

📢 News

• [2026-06-19] 🚀 Paper is now available on arXiv: arXiv:2606.19687.
• [2026-06-18] 🛰️ arXiv submission completed.
• [2026-06-05] 🎉 Paper accepted to the 1st Workshop on Robot Meets GNSS and Ranging for Seamless Autonomy, IEEE ICRA 2026.
• [2026-06] 🛠️ Code release is being cleaned for public use. See the roadmap below.

---

🌐 Community & Coverage

We appreciate independent discussions and community attention around this work. The links below are third-party materials and are not official implementations unless explicitly stated.

• 🔁 Independent reproduction — a WeChat article reproduced and discussed the main WheelRoute pipeline and results. Read here
• 📝 Technical review — Moonlight published an AI-generated summary and review of the arXiv paper. Read here
• 💬 Social mention — LinkedIn community post about the paper. View post

---

🎯 TL;DR

A wheeled robot enters a tunnel, GNSS disappears, and MEMS/odometry dead reckoning starts drifting. WheelRoute keeps the robot localized by matching the recent dead-reckoning trajectory to route-level map geometry, turning the route into a pseudo-position measurement inside an EKF fusion framework.

It is designed for practical UGV navigation systems: MEMS/GNSS/odometry in, route-constrained robust localization out.

---

✨ Highlights

• 🛞 Wheeled-robot degraded localization — targets UGV localization under GNSS interruption, blockage, and degraded reception.
• 🗺️ Route-level map prior — uses mission-route geometry as a lightweight constraint for drift suppression.
• 🧭 Trajectory-to-route matching — aligns recent dead-reckoning history with local route segments through 2D rigid registration.
• 🧩 Map-as-measurement fusion — converts route matching into EKF-compatible pseudo-position observations.
• 🛡️ Robust engineering design — trigger control, matching validation, route offset compensation, and correction limiting.
• 🚇 Real-world tunnel validation — long-tunnel, multi-segment tunnel, and curved-tunnel scenarios.

---

📅 Release Roadmap

Public release status. More details will be added as each component is ready.

• ICRA 2026 workshop acceptance
• arXiv submission completed
• arXiv identifier available
• Project page and video demo
• Reproducible code package
• Dataset access instructions

---

🧭 Method

The pipeline turns route geometry into a usable localization measurement.

1. Dead reckoning (DR) — MEMS IMU and odometry propagate a short-term trajectory when GNSS is degraded.
2. Route registration — the recent DR trajectory is aligned to local mission-route segments.
3. Pseudo-position observation — the matched route-referenced position becomes a measurement.
4. Robust EKF fusion — pseudo-position observations suppress accumulated drift while preserving continuity.

---

📊 Results

Evaluated on three real-vehicle tunnel scenarios under GNSS-degraded conditions. Position metrics are route-relative deviations in meters; heading metrics are in degrees.

Scenario	Method	Max Pos. Dev.	Mean Pos. Dev.	Pos. RMSE	Mean Heading Dev.	Heading RMSE
Long tunnel	Baseline	386.3	142.909	186.821	2.073	3.524
Long tunnel	WheelRoute	22.7	0.745	1.672	0.048	1.257
Curved tunnel	Baseline	32.6	7.908	10.376	-0.216	0.793
Curved tunnel	WheelRoute	27.5	1.431	1.889	-0.179	0.946
Multi-segment tunnel	Baseline	23.3	3.576	6.502	0.011	1.839
Multi-segment tunnel	WheelRoute	8.5	1.755	2.351	0.018	1.332

---

🚀 Getting Started

Code release in progress. The internal implementation has reproduced the tunnel experiments reported in the paper; the public interface and documentation are being cleaned.

git clone https://github.com/Fuu1718121/WheelRoute.git
cd WheelRoute

The first public code release will include an example sequence interface, configuration files, and plotting scripts for reproducing the reported localization metrics.

---

📦 Dataset

The experiments cover three real-vehicle GNSS-degraded tunnel scenarios:

• 🚇 Long tunnel
• 🧱 Multi-segment tunnel
• 🌀 Curved tunnel

Dataset release or access instructions will be added after cleanup.

---

📝 Citation

If you find this work useful, please consider citing:

@inproceedings{cui2026wheelroute,
  title     = {Route-Constrained Robust Fusion Estimation for MEMS/GNSS Integrated Navigation of Unmanned Ground Vehicles in GNSS-Degraded Environments},
  author    = {Cui, Jingzhi and Zhang, Chao and Mao, Yuliang and L{\"u}, Shaolin and Li, Dongmei and Che, Huan and Zhang, Rong},
  booktitle = {1st Workshop on Robot Meets GNSS and Ranging for Seamless Autonomy, IEEE ICRA},
  address   = {Vienna, Austria},
  date      = {2026-06-05},
  year      = {2026},
  eprint    = {2606.19687},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}

---

🙏 Acknowledgements

This work was conducted by Tsinghua University and Xiaomi Inc. We thank the Robot Meets GNSS and Ranging workshop organizers for hosting the accepted paper at IEEE ICRA 2026.

_{State Key Laboratory of Precision Space-time Information Sensing Technology, Tsinghua University}