TL;DR
Data-driven lateral velocity estimators often "collapse" when moved from their training vehicle to a structurally different one. This paper demonstrates that by carefully pruning input signals and applying physically motivated data augmentations—like enforcing left-right symmetry—neural networks can achieve robust zero-shot transfer, matching the performance of traditional model-based filters on unseen vehicles.
What Problem Does This Paper Address?
In the world of vehicle dynamics, lateral velocity ($v_y$) is a "holy grail" state. It’s critical for safety functions like Electronic Stability Control (ESC) and autonomous evasive steering, yet it cannot be measured directly with standard production sensors. While high-precision GNSS-aided inertial systems can measure it, they are far too expensive for series production.
Engineers typically turn to two solutions: model-based estimators (like Kalman filters) which rely on physical equations, or data-driven estimators (neural networks) which learn the mapping from available CAN bus signals. While neural networks are often more accurate on the specific car they were trained on, they are notorious for failing when deployed on a different vehicle—a problem known as the "cross-vehicle gap." This paper quantifies this gap for the first time on production vehicles and proposes a way to bridge it without needing any data from the new vehicle.
Methods
The authors used a Gated Recurrent Unit (GRU) network as their primary data-driven estimator. To improve its ability to generalize to unseen vehicles (zero-shot transfer), they focused on two "data-level" strategies rather than changing the network architecture:
- Systematic Input Selection: They tested six different sets of production CAN signals. The goal was to find a "lean" set that provides enough information for estimation but avoids vehicle-specific signals (like engine torque) that might lead to overfitting.
- Physically Motivated Data Augmentation: They introduced three augmentations during training to bake physical priors into the model:
- Symmetry Data Augmentation (SDA): Exploits the fact that cars are (mostly) mirror-symmetric. It flips the signs of lateral signals and swaps left/right wheel speeds.
- Channel Time Shift: Emulates varying CAN bus latencies by randomly delaying input channels.
- IMU Bias: Adds random constant offsets to accelerometer signals to simulate sensor drift and installation tolerances.
They compared their results against a classical Extended Kalman Filter (EKF) tuned on a dynamic single-track model.
Datasets
The study utilized real-world test drive data from three distinct series-production vehicles:
- Vehicle A (Compact Car): The sole training vehicle.
- Vehicle B (SUV): An unseen vehicle for zero-shot evaluation.
- Vehicle C (Van): Another unseen vehicle for zero-shot evaluation.
All vehicles were equipped with a high-end ADMA-G-ECO+ reference system to provide ground-truth lateral velocity. The datasets included dynamic handling maneuvers like circles and free-form driving.
Key Insights and Ideas
- The "Less is More" Principle for Transfer: Including drive and brake torques actually hurt transferability. Why? Because the network learns the specific relationship between torque and acceleration for a car of a certain mass. When moved to a heavier SUV, that relationship breaks. A robust 5-signal set (yaw rate, longitudinal/lateral acceleration, steering angle, and averaged wheel speed) proved best for transfer.
- Symmetry as a Strong Prior: Enforcing left-right symmetry (SDA) was the most effective augmentation. It helps the model ignore the inherent "left-turn bias" often found in specific test track recordings and forces it to learn a more general physical mapping.
- Quantifying the Gap: Without these interventions, the error on unseen vehicles was an order of magnitude higher than on the training vehicle. The "vehicle gap" (different mass/wheelbase) and "measurement gap" (different sensor biases/latencies) combine to create a significant hurdle for deep learning in automotive control.
Key Results
- Error Reduction: Combining Symmetry and IMU Bias augmentations reduced the Root Mean Square Error (RMSE) on unseen validation maneuvers by 13% for the SUV and 15% for the van.
- Matching the Baseline: On the SUV, the augmented GRU achieved an accuracy comparable to the model-based EKF, challenging the common assumption that data-driven models always fail in zero-shot scenarios.
- Worst-Case Performance: The augmentations significantly lowered the maximum absolute error across all vehicles, which is a critical metric for safety-relevant systems.
Limitations and Future Work
The authors acknowledge several limitations:
- Single Training Source: The model was only trained on one vehicle. Training on a more diverse fleet would likely further narrow the transfer gap.
- Extrapolation Limits: On the van (Vehicle C), the EKF still outperformed the GRU. This is because the van reached lateral velocities outside the range seen in the compact car's training data. Neural networks still struggle with this kind of "out-of-range" extrapolation where physical models excel.
- Future Directions: Promising next steps include combining these data-level measures with architectural priors (physics-informed neural networks) and exploring online adaptation where the model continues to learn once deployed on a new vehicle.
Summary updated