WING: Wheel-Inertial Neural Odometry with Ground Manifold Constraints

TL;DR

WING tackles robustness of pose estimation for ground robots when exteroceptive sensing is unreliable by employing interoceptive data processed through neural networks that correct IMU and wheel-encoder measurements and enable time-varying uncertainty estimation. It integrates a globally continuous dual cubic B-spline ground manifold as soft constraints and fuses all information in a space-based sliding-window EKF, preserving the kinematic model while learning corrections. The key contributions are the IMU De-Bias Net, Wheel Encoder Net with learned covariances, and the dual cubic B-spline manifold that provides global soft constraints; experiments on KAIST Urban and NCLT datasets show improved translation and rotation RMSE compared to state-of-the-art interoceptive-only methods. The approach reduces reliance on exteroceptive sensors and runs in real time on common hardware, offering a robust navigation input for degraded sensing conditions.

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.

Figures (7)

• Figure 1: The pipeline of our proposed odometry. Our proposed odometry estimates pose by processing the interoceptive sensor data through neural networks while simultaneously reconstructing a continuous cubic B-spline manifold. The middle row of the figure shows the estimated position and terrain of scenario urban17 in the KAIST Urban Dataset jeong2019complex, where A and B represent the corner and straight routes, respectively.
• Figure 2: The demonstration of space-based sliding-window update strategy while the vehicle is moving in the rightward direction. Yellow represents active control points and patches of the manifold, blue represents static control points and patches, and gray represents marginalized control points and patches.
• Figure 3: Demonstration of different algorithms for pose estimation. From left to right, there are 3D estimated trajectories of urban09, urban15, and urban17 of the KAIST Urban Dataset jeong2019complex.
• Figure 4: Predicted time-varying bias estimation and its corresponding covariance from IMU De-Bias Net on the urban11 of the KAIST Urban Dataset jeong2019complex.
• Figure 5: Standard deviation $\sigma$ against the errors in ${\boldsymbol{\widehat{v}}}_{i,i+n}$ (first row) and ${{\boldsymbol q}}_{i,i+n}$ (second row) along the x, y, and z axes. Error $dv$ represents $\boldsymbol {v}_{i,i+n} - {\boldsymbol{\widehat{v}}}_{i,i+n}$, and error $dq$ denotes $\mathrm{Log} ( \left({\boldsymbol{\widehat{q}}}_{i,i+n} \right)^* \otimes {\boldsymbol q}_{i,i+n} )$. The dashed red line represents the 3$\sigma$ region and the red text indicates the ratio of error points that fall within this 3$\sigma$ region.