Cross-Vehicle Transferability of Lateral Velocity Estimation on Production Vehicles through Input Selection and Data Augmentation
Authors: Hendrik Schäfke, Simon F. G. Ehlers, Thomas Seel, Askar Vagapov, Christoph Schweers, Ahmed Trabelsi, Daniel O. M. Weber
Organizations: Institute of Mechatronic Systems, Leibniz University Hannover, 30823, Garbsen, Germany · L3S Research Center, Leibniz University Hannover, 30167, Hannover, Germany · IAV GmbH, 38518, Gifhorn, Germany
Lateral velocity is safety-critical for vehicle-dynamics control but cannot be measured with production sensors. Model-based estimators such as Kalman filters are established, whereas data-driven estimators are more accurate within the training regime but suspected to lose accuracy beyond it. Their transfer to a structurally different vehicle has rarely been tested. We record real-world test drives of three series-production vehicles, train a neural-network estimator on production CAN signals of a compact car, and evaluate it zero-shot on an SUV and a van against a GNSS-aided inertial reference. On the unseen vehicles, the network's error exceeds that on the training vehicle by an order of magnitude, a gap not quantified on production vehicles before. A systematic comparison of input sets from ten production CAN signals identifies a five-signal set that transfers most robustly. To narrow this gap, we introduce three physically motivated augmentations of training data. Two improve generalization, namely exploiting the vehicle's left-right symmetry and emulating accelerometer bias. Combined, they reduce the error on the validation maneuvers of the SUV and van by 13% and 15%. Contrary to common expectation, on the SUV the data-driven estimator trained with both augmentations matches a model-based estimator tuned on the same training data.
Figures & tables
Study
Target
Production vehicle
Cross-vehicle
Production CAN only
Data augmentation
Input-set selection
Melzi and Sabbioni [11]
β
✓
–
✓
✓
–
Hwang et al. [18]
vy
–
–
✓
–
✓
Tristano et al. [19]
β
( ∼ )
–
✓
–
✓
Ayyad et al. [22]
vx , vy
( ∼ )
( ∼ )
–
–
–
Da Lio et al. [21]
vy
( ∼ )
✓
✓
–
–
James et al. [6]
vx
✓
( ∼ )
( ∼ )
–
–
Table 1: Positioning against prior neural-network lateral-velocity studies. Of the 18 surveyed studies, the table lists those that address at least one of cross-vehicle evaluation, data augmentation, and systematic input selection. The remaining surveyed studies address none of the three. James et al. [6] model the longitudinal dynamics and Kontos et al. [23] forecast the yaw rate, and both are listed in addition because they address the cross-vehicle question.
Figure 1: Overview of the study on cross-vehicle zero-shot estimation of the lateral velocity vy . (a) A GRU estimator is trained on production-CAN data of Vehicle A and transferred zero-shot to the structurally different, unseen Vehicle B and Vehicle C. All three vehicles are drawn generically for confidentiality, with illustrations generated by Google Gemini 3.1 Pro. (b) The estimation pipeline with the four measures. The input-set selection acts on the input signals, and the three data augmentations act on the training windows only. (c) Estimated vy during a cornering maneuver on the unseen Vehicle B, where augmentation restores the tracking of the ADMA reference that the GRU without augmentation loses.
Vehicle A
Vehicle B
Vehicle C
Role
training
unseen
unseen
Body type
compact car
SUV
van
Drivetrain
plug-in hybrid
battery-electric
battery-electric
Driven axles
front
rear
front and rear
Mass in kg
1720
2150
2650
Wheelbase in m
2.62
2.8
3.0
Table 2: Main characteristics of the three vehicles. The values of Vehicle B and Vehicle C are rounded.
Figure 2: Dataset distributions of the three vehicles over all splits: (a) lateral acceleration ay , (b) yaw rate ψ˙ , and (c) longitudinal velocity vx , each against the lateral velocity vy (ADMA reference). Contour lines show the sample density per vehicle (one color and line style each). The faint points are individual samples and preserve the distribution tails. The same vy arises at different combinations of ay and ψ˙ on the three vehicles, reflecting their different dynamics (mainly the differing masses), and the campaigns cover different vx ranges. These two differences correspond to the vehicle gap and the measurement gap.
Set
n inputs
Wheel speeds
IMU
Steering
Torques
Closest prior work
ax
ay
ψ˙
δ
Mdrive
Mbrake
1
10
four (per wheel)
✓
✓
✓
✓
✓
✓
[ 7 , 4 ] (approx.)
2
8
four (per wheel)
✓
✓
✓
✓
–
–
[ 9 , 3 , 12 , 13 , 5 ] (approx.)
3
6
two (per side)
✓
✓
✓
✓
–
–
none
4
5
one (average)
✓
✓
✓
✓
–
–
[ 8 , 14 , 11 , 21 ] (approx.)
5
4
none
✓
✓
✓
✓
–
–
none
Table 3: The six input sets as a signal matrix, with n the number of input signals. Every set contains the IMU core ax , ay , ψ˙ , so the sets differ only in the wheel-speed resolution, the steering angle, and the drive and brake torques. Wheel speeds enter per wheel ( nFL,nFR,nRL,nRR ), per side as left and right averages ( nL,nR ), or as a single all-wheel average ( navg ). The last column names the published neural-network vy estimator with the most similar input set, where (approx.) marks a close but not exact match, and none a set not used in the literature so far.
Algorithm 1 Data augmentations applied to the training data of Vehicle A. (a) Symmetry Data Augmentation, (b) Channel Time Shift, and (c) IMU Bias, each shown with an example of its effect on a signal. S input sequence with channels sc , vy target, p application probability, αmax maximum shift in samples, σ bias standard deviation in m/s2 on the unnormalized acceleration signals.
Figure 3: Validation RMSE of vy across the six input sets (Experiment 1), one panel per vehicle. Boxes span the interquartile range of the nine seeds with the median as the center line, whiskers extend to 1.5 times the interquartile range, and the open circle marks the per-set mean. The full set (Set 1) and the IMU-only set (Set 6) degrade most on the unseen vehicles, whereas the mid-sized sets transfer best.
Figure 4: Augmentation ablation on Set 4 (Experiment 2), one panel per vehicle, with the GRU without augmentation next to the three augmentations applied individually. Each box aggregates the nine seeds of one condition and is drawn as in Figure 3 , with the seeds added as dots. Symmetry Data Augmentation lowers the out-of-domain error on both unseen vehicles and IMU Bias on Vehicle C, whereas Channel Time Shift does not, which motivates its exclusion from the combined augmentation.
Validation maneuvers
Test maneuvers
Vehicle
Estimator
RMSE
Max. error
RMSE
Max. error
Vehicle A
GRU (no aug.)
0.025 (0.002)
0.21
0.015 (0.002)
0.10
GRU (augmented)
0.022 (0.001)
0.11
0.016 (0.003)
0.08
EKF
0.030
0.12
0.035
0.09
Vehicle B
GRU (no aug.)
0.239 (0.011)
0.97
0.180 (0.008)
0.89
GRU (augmented)
0.208 (0.003)
0.86
0.173 (0.009)
0.79
Table 4: RMSE and maximum absolute error of vy in m/s in Experiment 3 for the GRU without and with the combined augmentation and for the EKF tuned on Vehicle A. GRU values are means over nine seeds, with the standard deviation of the RMSE in parentheses, and the maximum error is that of the worst maneuver. The input set and the augmentations are selected on the validation maneuvers of Vehicle C, and the test maneuvers take no part in any selection. Bold marks the lowest value per vehicle and column.
Figure 5: Cross-vehicle RMSE of vy , grouped per vehicle. Boxes show the GRU without and with the combined augmentation (boxes and whiskers as in Figure 3 , dots for the individual seeds), and the black star marks the model-based EKF tuned on Vehicle A. On these validation maneuvers, the augmented GRU reaches a lower error than the EKF on Vehicle B, and the EKF reaches a lower error on Vehicle C.
Figure 6: Qualitative cross-vehicle prediction on two validation maneuvers of Vehicle B, (a) dynamic circles and (b) free handling, and on the test maneuver of Vehicle C, (c) circles. Top: longitudinal velocity vx of the ADMA reference for maneuver context. Middle: lateral velocity vy of the ADMA reference, the GRU without and with the combined augmentation, and the EKF, all trained or tuned on Vehicle A. Bottom: residuals with the maneuver RMSE of each estimator.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Group
Parameter
Value
Measured
mass m
1720kg
yaw inertia Iz
2066kgm2
front axle distance lf
1.16m
rear axle distance lr
1.46m
Tuned
front cornering stiffness Cf
140510N/rad
rear cornering stiffness Cr
177693N/rad
Appendix
Table A.1: EKF parameters, measured on Vehicle A or tuned on its training maneuvers. All parameters are applied unchanged to Vehicle B and Vehicle C. The noise covariances are listed as standard deviations, with Q=diag(qvy2,qψ˙2) and R=diag(ray2,rψ˙2) .
Set
Hidden
Learning
Input
Weight
Window
size
rate
dropout
decay
(samples)
1
86
7.2\times10−4
2.0\times10−2
3.5\times10−5
200
2
124
3.8\times10−4
4.3\times10−3
3.1\times10−5
200
3
124
9.6\times10−4
3.7\times10−2
4.8\times10−4
200
4
86
3.3\times10−4
3.8\times10−3
7.8\times10−5
250
5
95
4.5\times10−4
4.1\times10−4
1.1\times10−3
150
Appendix
Table A.2: ASHA-selected GRU configuration per input set (Stage 0), trained on Vehicle A. Fixed across all sets: single-layer GRU, AdamW optimizer, 100 epochs, MAE loss, batch size 32 , recurrent dropout and auxiliary losses off.
Figure B.1: Component-wise ablation of Symmetry Data Augmentation on input Set 2, which carries the four individual wheel speeds so that the left/right swap is active, with one panel per vehicle. The three conditions per panel are the GRU without augmentation from Experiment 1, the sign flip alone ( flip-only ), and the sign flip combined with the wheel swap ( full ), and each box aggregates the nine seeds with the individual seeds as dots. Most of the gain is due to the sign flip, while the wheel swap adds a small further cross-vehicle reduction.
Cross-domain generalization studies how machine-learning models retain reliable performance when data, sensors, populations, or tasks differ from training.
College of Automotive Engineering, Jilin University, Changchun, China. · School of Vehicle and Mobility, Tsinghua University, Beijing, China. · College of Mechanical Engineering, Tianjin University of Science and Technology, Tianjin, China.