Physics-Informed Neural State-Space Modeling of Battery-Electric Vehicle Dynamics for Closed-Loop Automated Parking Simulation

Jul 16, 2026

Upload by Aurora

Physics-informed neural state-space modeling is presented for the parking regime of a production battery-electric sedan, fully identified from field-test maneuvers. At low parking speeds, the model includes effects neglected by kinematic idealization: actuator lag, drivetrain creep, brake-hold transitions through standstill, and frequent motion-direction reversals. Training adds a gear-conditioned velocity constraint and learns yaw-rate residuals over a kinematic bicycle prior, shifting model capacity to physical deviations. Dedicated actuator submodels reproduce commanded-to-actual drive, brake, and steering behavior, enabling real-time model-in-the-loop simulation. Identified from only 16 tests, it generalizes to held-out maneuvers and achieves Good ratings under ISO/TS 18571 state objectives, supporting virtual pre-calibration of automated-parking planning and control.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

Introduction

Modeling gaps in the parking regime

Closed-loop fidelity versus open-loop accuracy

Data-driven and physics-informed modeling context

Method overview

Physics-Informed Neural State-Space Modeling of Battery-Electric Vehicle Dynamics for Closed-Loop Automated Parking Simulation Sirong Pan, Guannan Tian, and Pan Song arXiv :2607 .03000v1 [ ee ss . SY] 3 Jul 2026 Abstract—This paper contributes to vehicle dynamics modeling by introducing a physics-informed neural state-space model tailored for the parking regime of a production battery-electric sedan, identified entirely from field-test maneuvers. At parking speeds the model captures what the kinematic idealization omits, including actuator lag, drivetrain creep, brake-hold transitions through standstill, and frequent reversals of the motion direction. A gear-conditioned velocity constraint is imposed during training, and the yaw rate is read out as a learned residual on a kinematicbicycle prior, so that the network devotes its capacity to the deviation from physics rather than to its reproduction. These training-time physics make the customary inference-time state limiter redundant. The commanded-to-actual behavior of the drive, brake, and steering actuators is reproduced by dedicated submodels, for which signal fidelity proves an unreliable proxy for closed-loop value; tuning the brake on its velocity consequence rather than on its own signal reverses the verdict reached at the signal level. The model generalizes to held-out maneuvers in fully open-loop simulation, and, despite being identified from only 16 field tests, the assembled command-to-vehicle chain earns Good ratings on the vehicle states under the ISO/TS 18571 objective rating metric. Embedded as the real-time plant of an interactive simulator, it enables a production-representative planning stack to park the vehicle through the learned dynamics. This makes the model suitable for pre-calibrating an automated-parking planning and control stack in the virtual development phase without the manufacturer’s proprietary chassis and actuator parameters. Index Terms—Automated parking, electric vehicles, model-inthe-loop simulation, neural state-space models, physics-informed neural networks, system identification, vehicle dynamics. I. INTRODUCTION Accurate vehicle dynamics modeling is fundamental to the development, tuning, and evaluation of automated driving functions, including automated parking. The parking regime, however, is poorly served by established modeling practice. At speeds below walking pace the classical dynamics models collapse to kinematics, yet the real behavior is dominated by the effects those kinematics discard: actuator lag and Manuscript received XX XX, 2026 . This work was supported by the Wuhu Intelligent Logistics Technology Research and Development Center under Grant WHSYFZX202402 . (Corresponding author: Pan Song.) S. Pan is with the Modern Logistics and Intelligent Manufacturing College, Wuhu Vocational Technical University, Wuhu 241003, Anhui, China (e-mail: [900289@whit.edu.cn](900289@whit.edu.cn)) . G. Tian is with the Department of Vehicle Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China (e-mail: tianguan[nan@nuaa.edu.cn](nan@nuaa.edu.cn)) . P. Song is with the Kaiyang Laboratory, Chery Automobile Co., Ltd., Wuhu 241009, Anhui, China (e-mail: [songpan14@gmail.com](songpan14@gmail.com); ORCID: 0000-0003- 0814-6824) . The source code and trained models are available at [https://github.com/](https://github.com/)[ ](https://github.com/)pansong/PyNSSM-Parking, and the parking simulator at [https://github.com/](https://github.com/)[ ](https://github.com/)[pansong/auto-parking-sim.](pansong/auto-parking-sim.) hysteresis, the creep torque of an electric driveline, brakehold transitions through standstill, and frequent reversals of the motion direction. Production practice nonetheless adopts the kinematic-bicycle model for planning, where it is appropriate, and for execution-level simulation, where it idealizes away the phenomena that decide whether a maneuver lands in the bay. Data-driv

FAQs

How does the physics-informed neural state-space model improve parking dynamics modeling compared with kinematic bicycle models?

It captures actuator lag, drivetrain creep, brake-hold transitions through standstill, and frequent reversals that kinematic idealization omits. It also models yaw-rate as a learned residual over a kinematic prior, focusing capacity on deviations from physical behavior.

What constraints and learning targets are used during training?

A gear-conditioned velocity constraint is imposed during training. The model reads out yaw rate as a learned residual on top of a kinematic-bicycle prior.

Why is closed-loop evaluation emphasized for automated parking?

A model that is accurate one step ahead can still fail as a closed-loop plant because errors compound and the controller reacts to the model’s own outputs. For parking, closed-loop fidelity determines whether the learned dynamics can support real-time planning and control.

Jul 16, 2026

Upload by Aurora