Uncertainty-Aware Hybrid Machine Learning in Virtual Sensors for Vehicle Sideslip Angle Estimation

Abinav Kalyanasundaram, Karthikeyan Chandra Sekaran, Philipp Stäuber, Michael Lange, Wolfgang Utschick, Michael Botsch · 2025

Read Uncertainty-Aware Hybrid Machine Learning in Virtual Sensors for Vehicle Sideslip Angle Estimation by Abinav Kalyanasundaram, Karthikeyan Chandra Sekaran, Philipp Stäuber, Michael Lange, Wolfgang Utschick, Michael Botsch (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-06105.txt.

Introduction

How sideways a car slides in a turn decides whether its safety systems react in time, yet measuring that angle needs pricey optical gear. Abinav Kalyanasundaram and co-authors build a virtual sensor instead in a 2025 paper, marrying a learning model with the physics of vehicle motion. Each side states how unsure it is, and the blend leans on the surer voice.

The team also releases a real-world driving data set from advanced sensors to test such methods. Their hybrid beat the parts taken alone on that data. The design aims at cheaper cars that still know exactly how they are moving.

Catalog record

CREATORAbinav Kalyanasundaram, Karthikeyan Chandra Sekaran, Philipp Stäuber, Michael Lange, Wolfgang Utschick, Michael Botsch
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-06105

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Uncertainty-Aware Hybrid Machine Learning in Virtual Sensors for Vehicle Sideslip Angle Estimation Abinav Kalyanasundaram, Karthikeyan Chandra Sekaran, Philipp Stäuber, Michael Lange, Wolfgang Utschick, Michael Botsch Abstract Precise vehicle state estimation is crucial for safe and reliable autonomous driving. The number of measurable states and their precision offered by the onboard vehicle sensor system are often constrained by cost. For instance, measuring critical quantities such as the Vehicle Sideslip Angle (VSA) poses significant commercial challenges using current optical sensors. This paper addresses these limitations by focusing on the development of high-performance virtual sensors to enhance vehicle state estimation for active safety. The proposed Uncertainty-Aware Hybrid Learning (UAHL) architecture integrates a machine learning model with vehicle motion models to estimate VSA directly from onboard sensor data. A key aspect of the UAHL architecture is its focus on uncertainty quantification for individual model estimates and hybrid fusion. These mechanisms enable the dynamic weighting of uncertainty-aware predictions from machine learning and vehicle motion models to produce accurate and reliable hybrid VSA estimates. This work also presents a novel dataset named Real-world Vehicle State Estimation Dataset (ReV-StED) , comprising synchronized measurements from advanced vehicle dynamic sensors. The experimental results demonstrate the superior performance of the proposed method for VSA estimation, highlighting UAHL as a promising architecture for advancing virtual sensors and enhancing active safety in autonomous vehicles. I Introduction Safe autonomous driving and Advanced Driver Assistance Systems (ADAS) require an accurate estimation for the ego vehicle state. The knowledge about the dynamic states of a vehicle acts as a priori for downstream tasks such as autonomous emergency braking, electronic stability control, and path planning. Among these dynamic states, the VSA denoted as \beta , is a key indicator of vehicle stability and rider comfort making it crucial for control strategies in active safety systems [ 1 ] . However, sensors that directly measure VSA for real-time monitoring are expensive to implement in production cars [ 2 ] . Alternatively, virtual sensors can estimate vehicle dynamic states by leveraging data from available sensors without space and cost constraints. The VSA is often estimated through model-based virtual sensors employing different versions of the Kalman Filter (KF) . The recent trend towards software-defined vehicles has paved way for the adoption of Machine Learning (ML) -based virtual sensors [ 3 ] . This work presents a virtual sensor designed for reliable VSA estimation from onboard sensor data. State-of-the-art (SOTA) methods in virtual sensors for VSA estimation use observer-based algorithms [ 4 , 5 ] or ML -based approaches [ 1 , 2 ] . Observer-based methods struggle to capture the non-linearity in vehicle dynamics and their performance relies on external correction data [ 6 ] . While ML -based approaches can handle non-linearities, they suffer from the “long-tail” problem, where more erroneous predictions occur in underrepresented scenarios. They also typically lack the ability to quantify the uncertainty in their predictions [ 7 ] . In safety-critical applications, such as autonomous driving, understanding these uncertainties is essential for ensuring reliable estimates from virtual sensor outputs [ 8 ] . Fig. 1 : Overview of the proposed UAHL architecture for virtual VSA sensor. This work aims to address these limitations by introducing the UAHL architecture for virtual sensors. It consists of a ML model in parallel with vehicle motion models as shown in Fig. 1 . The ML branch considers the VSA estimation as a time series forecasting task from onboard sensor data, while the model-based approaches incorporate vehicle dynamics knowledge into the architecture to reduce errors in “long-tail” scenarios. Uncertainty quantification is implemented independently in each branch to leverage their complementary strengths during hybrid fusion. This uncertainty-aware hybrid fusion improves the performance and reliability of estimates from the virtual sensor. The main contributions of this work are as follows: • Introduction of a novel UAHL architecture for a reliable virtual VSA sensor, with the following components: a) A ML model to estimate VSA and its uncertainty based on the Informer [ 9 ] . b) Vehicle Motion Models (VMMs) with uncertainty quantification for its VSA estimates. c) Hybrid fusion strategies to combine ML and VMMs . • Publication of a novel dataset named ReV-StED to support research in virtual sensors. 1 1 1 Dataset and code to replicate the results will be made open-source upon acceptance. A dataset sample is provided for reference in https://github.com/MB-Team-THI/UAHL-RevStED • Evaluation of the performance and reliability of UAHL architecture as a virtual sensor for VSA estimation. II Related works Virtual sensors are an active area of research in intelligent vehicles, enabling advancements across a variety of domains, including vehicle dynamics [ 10 ] , battery thermal management [ 11 ] , electric drivetrain [ 12 ] , etc. They have huge potential in VSA estimation, defined as the angle \beta between a vehicle’s longitudinal axis and the direction of its velocity vector at the Center of Gravity (CoG) [ 2 ] . The existing research works on virtual estimation of the VSA can be categorized into observer-based approaches and data-driven ML methods [ 13 ] . The observer-based approaches like the Luenberger observer, the sliding mode observer and the Kalman filter utilize vehicle models for state estimation [ 4 ] . They are then combined with measurements from Global Positioning Systems (GPS) or inertial measurement units for VSA estimation. The vehicle models can be either kinematic [ 14 , 15 ] or dynamic [ 16 ] . While kinematic models offer simplicity from complex parameters, they can suffer in low yaw rate conditions as the system becomes unobservable [ 17 ] . Dynamic models incorporate forces and moments acting on the vehicle and vary in complexity depending on the type of model used, such as single-track [ 16 ] , two-track, and tire models [ 18 ] . Although dynamic models can provide good VSA estimation, their accuracy heavily depends on reliable parameters identification, the tire model and the tire wear [ 19 ] . Recently [ 5 ] attempted to address these challenges using an interacting multiple model KF with GPS data. Many observer-based approaches depend on external correction data [ 4 ] , which can be challenging to acquire. Moreover, their high accuracy is generally limited to specific vehicle states where the models are most effective [ 5 ] . The ML -based approaches use onboard [ 2 ] or external sensor data to directly estimate VSA . Dynamic Neural Networks (DNN) [ 1 ] , Recurrent Neural Networks [ 2 ] and Long Short-Term Memory (LSTM) models [ 20 ] are some of the promising architectures. They exhibited similar performance levels as observer-based approaches in certain maneuvers [ 21 ] . The main challenge with ML -based estimation is their inability to quantify uncertainty and performance reliance on dataset quality. In this work, some limitations of both observer-based and ML -based approaches are addressed, while leveraging their individual strengths, through the proposed UAHL architecture for virtual sensors. III methodology This section describes the problem formulation, preliminaries and the proposed UAHL architecture. The overall architecture is shown in Fig. 2 and a detailed explanation about each UAHL component including its integration is provided in the subsequent subsections. III-A Problem Formulation Let the

Read the complete text