TL;DR
WING (Wheel-Inertial Neural Odometry) is a robust localization system for ground robots that combines deep learning with a globally continuous $C^2$ ground manifold model. By using neural networks to correct sensor noise and a dual cubic B-spline to enforce smooth terrain constraints, WING achieves state-of-the-art accuracy in challenging environments where GPS, cameras, and LiDAR often fail.
What Problem Does This Paper Address?
Reliable localization is the "holy grail" for autonomous ground robots. While most systems rely on exteroceptive sensors like LiDAR, cameras, or GPS, these sensors are fragile: cameras fail in the dark, GPS disappears indoors, and LiDAR struggles in dusty or featureless tunnels.
In these "degraded" scenarios, robots must rely on interoceptive sensors—Inertial Measurement Units (IMUs) and wheel encoders. However, these are notoriously "drift-prone." IMUs suffer from accumulating bias, and wheel encoders provide inaccurate velocity when wheels skid on gravel or bounce over bumps. Previous attempts to fix this either treated the robot's motion as a "black box" (ignoring physics) or assumed the ground was a series of flat, disconnected tiles, which fails to capture the smooth, continuous nature of real-world terrain.
Methods
The authors propose a hybrid framework that blends the strengths of deep learning with classical state estimation:
- Neural Sensor Correction: Instead of replacing the kinematic model with a neural network, WING uses two specialized networks—IMU De-Bias Net and Wheel Encoder Net. These networks process raw data to output corrected measurements and, crucially, time-varying uncertainty (covariance). This allows the system to "know" when to distrust a sensor (e.g., during a skid).
- Global Ground Manifold: The core innovation is modeling the environment as a dual cubic B-spline manifold. Unlike "segmented" models that treat the ground as independent patches, this model ensures $C^2$ continuity—meaning the terrain is smooth in both its slope and its curvature. This provides a powerful "soft constraint" on the robot’s position and tilt.
- Space-Based Sliding Window Filter: The system uses an Extended Kalman Filter (EKF) with a "space-based" window. As the robot moves, it populates a grid of control points for the B-spline. This allows the robot to "remember" the terrain it just traversed, using historical spatial information to refine its current pose estimate.
- Yaw-Independent Convention: To prevent yaw drift from corrupting the tilt and velocity estimates, the authors utilize a specific attitude convention that decouples yaw from the gravity-aligned tilt components.
Datasets
The researchers evaluated WING on two major benchmarks:
- KAIST Urban Dataset: A large-scale dataset featuring complex urban environments in South Korea, captured with high-quality IMUs and wheel encoders.
- NCLT Dataset (University of Michigan): Used primarily to test generalization. The authors trained their networks on KAIST data and tested them on NCLT to prove the system works across different robot platforms and environments without retraining.
Key Insights and Ideas
- Physics-Informed Learning: By using NNs only to correct sensor biases rather than predicting the entire motion, the model retains the underlying physics of the robot. This makes it much more robust when moving from one dataset to another.
- The "Smoothness" Prior: Real ground isn't just a collection of planes; it’s a continuous surface. Enforcing $C^2$ continuity via B-splines acts as a "regularizer" for the trajectory, preventing the erratic vertical jumps often seen in pure inertial odometry.
- Uncertainty is Key: Most previous learning-based odometry used fixed noise parameters. WING’s ability to dynamically predict covariance means the filter can adapt in real-time to high-vibration or high-skid scenarios.
Key Results
WING significantly outperformed existing state-of-the-art (SOTA) methods like TLIO and RINS-W:
- Accuracy: On the KAIST urban07 sequence, WING achieved an Absolute Trajectory Error (ATE) of 8.10m, compared to 30.99m for a traditional EKF and over 90m for some other learning-based methods.
- Generalization: Even when trained on KAIST and tested on the NCLT dataset, WING maintained high accuracy (e.g., 8.22m ATE on NCLT 13/04/05), proving it isn't just "memorizing" specific routes.
- Efficiency: The system runs in real-time, with the entire fusion process taking only 0.9ms per update on a standard desktop setup.
- Ablation: The study showed that the "Continuous B-spline" model was the single most important factor in reducing rotation and translation errors compared to segmented or flat-ground assumptions.
Limitations and Future Work
- Static Ground Assumption: The current manifold model assumes the ground is static. It might struggle in environments with highly deformable surfaces (like deep sand) or moving platforms.
- Fixed Manifold Parameters: The size of the "knot vector" (the grid size for the B-spline) is currently pre-defined. The authors suggest that future versions could use a learning-based method to dynamically adjust the manifold resolution based on terrain complexity.
- One-Way Information Flow: Currently, the neural networks don't "know" about the ground manifold. A tighter coupling where ground parameters are fed back into the NNs could further improve sensor correction.
Summary updated