WING: Wheel-Inertial Neural Odometry with Ground Manifold Constraints

Chenxing Jiang, Kunyi Zhang, Sheng Yang, Shaojie Shen, Chao Xu, Fei Gao

Abstract

In this paper, we propose an interoceptive-only odometry system for ground robots with neural network processing and soft constraints based on the assumption of a globally continuous ground manifold. Exteroceptive sensors such as cameras, GPS and LiDAR may encounter difficulties in scenarios with poor illumination, indoor environments, dusty areas and straight tunnels. Therefore, improving the pose estimation accuracy only using interoceptive sensors is important to enhance the reliability of navigation system even in degrading scenarios mentioned above. However, interoceptive sensors like IMU and wheel encoders suffer from large drift due to noisy measurements. To overcome these challenges, the proposed system trains deep neural networks to correct the measurements from IMU and wheel encoders, while considering their uncertainty. Moreover, because ground robots can only travel on the ground, we model the ground surface as a globally continuous manifold using a dual cubic B-spline manifold to further improve the estimation accuracy by this soft constraint. A novel space-based sliding-window filtering framework is proposed to fully exploit the $C^2$ continuity of ground manifold soft constraints and fuse all the information from raw measurements and neural networks in a yaw-independent attitude convention. Extensive experiments demonstrate that our proposed approach can outperform state-of-the-art learning-based interoceptive-only odometry methods.

Introduction

Reliable localization is a fundamental requirement for autonomous ground robots. While exteroceptive sensors like LiDAR, cameras, and GPS are commonly used, they are susceptible to failure in "degraded" environments—such as dark indoor spaces, dusty tunnels, or featureless urban canyons. In these scenarios, robots must rely on interoceptive sensors: Inertial Measurement Units (IMUs) and wheel encoders. However, these sensors suffer from significant drift due to noise, bias accumulation, and physical phenomena like wheel slip.

WING (Wheel-Inertial Neural Odometry) addresses these challenges by combining deep learning for sensor correction with a globally continuous ground manifold model. This hybrid approach leverages the strengths of neural networks to handle complex sensor noise while enforcing physical consistency through terrain constraints.

Methodology

The WING framework consists of several key components designed to minimize drift and improve estimation accuracy.

Neural Sensor Correction

Instead of a "black-box" end-to-end motion model, WING employs two specialized networks: the IMU De-Bias Net and the Wheel Encoder Net. These networks process raw sensor buffers to output corrected measurements and time-varying uncertainty (covariance). This allows the system to dynamically weight sensor inputs based on their predicted reliability, such as discounting wheel encoders during high-slip events.

System overview of WING, integrating neural sensor correction with a global ground manifold constraint.

Global Ground Manifold

A core innovation of WING is modeling the environment as a dual cubic B-spline manifold. Unlike previous methods that assume a flat ground or independent local patches, this model ensures $C^2$ continuity, meaning the terrain is smooth in both slope and curvature. This provides a powerful "soft constraint" on the robot’s position and orientation.

Space-Based Sliding Window Filter

The system utilizes an Extended Kalman Filter (EKF) within a space-based sliding window framework. As the robot traverses the environment, it populates a grid of B-spline control points. This allows the filter to utilize historical spatial information to refine current pose estimates, effectively "remembering" the terrain to constrain future motion.

Experimental Results

The system was evaluated on the KAIST Urban Dataset and the NCLT Dataset to test both large-scale accuracy and generalization capabilities.

Accuracy and Generalization

WING significantly outperformed state-of-the-art (SOTA) methods such as TLIO and RINS-W. On the KAIST urban07 sequence, WING achieved an Absolute Trajectory Error (ATE) of 8.10m, a substantial improvement over traditional EKF (30.99m) and other learning-based approaches. Furthermore, networks trained on KAIST data generalized well to the NCLT dataset without retraining, maintaining high accuracy (e.g., 8.22m ATE on NCLT 13/04/05).

Trajectory comparison on large-scale datasets showing WING's superior performance over existing methods.

Efficiency

The entire fusion process is highly efficient, taking approximately 0.9ms per update on standard hardware, making it suitable for real-time robotic applications.

Conclusion

WING demonstrates that integrating physics-informed learning with continuous manifold constraints can dramatically improve the reliability of interoceptive-only odometry. By modeling the ground as a smooth, continuous surface and using neural networks to adaptively correct sensor noise, WING provides a robust localization solution for challenging environments where traditional exteroceptive sensors fail.

Summary updated

Abstract

In this paper, we propose an interoceptive-only odometry system for ground robots with neural network processing and soft constraints based on the assumption of a globally continuous ground manifold. Exteroceptive sensors such as cameras, GPS and LiDAR may encounter difficulties in scenarios with poor illumination, indoor environments, dusty areas and straight tunnels. Therefore, improving the pose estimation accuracy only using interoceptive sensors is important to enhance the reliability of navigation system even in degrading scenarios mentioned above. However, interoceptive sensors like IMU and wheel encoders suffer from large drift due to noisy measurements. To overcome these challenges, the proposed system trains deep neural networks to correct the measurements from IMU and wheel encoders, while considering their uncertainty. Moreover, because ground robots can only travel on the ground, we model the ground surface as a globally continuous manifold using a dual cubic B-spline manifold to further improve the estimation accuracy by this soft constraint. A novel space-based sliding-window filtering framework is proposed to fully exploit the $C^2$ continuity of ground manifold soft constraints and fuse all the information from raw measurements and neural networks in a yaw-independent attitude convention. Extensive experiments demonstrate that our proposed approach can outperform state-of-the-art learning-based interoceptive-only odometry methods.
BibTeX
@article{jiang2025wing,
  author = {Chenxing Jiang and Kunyi Zhang and Sheng Yang and Shaojie Shen and Chao Xu and Fei Gao},
  title = {{WING: Wheel-Inertial Neural Odometry with Ground Manifold Constraints}},
  journal = {IEEE Trans. Intell. Veh.},
  year = {2025},
  url = {https://openreview.net/forum?id=KAHIwGDeRX},
}
Figure