Development of a Neural Network-Based Autonomous Vehicle Simulator with an Actuator-Separated Model and Multi-Step Prediction Learning
Conventional vehicle dynamics simulations are often built based on explicit physical models. However, in real vehicles, the delay and response characteristics of steering, acceleration, and braking actuators can directly affect prediction errors. In this study, we propose a neural network-based simulation method that predicts changes in vehicle states using real driving data and actuator response characteristics, without constructing explicit vehicle dynamics equations. The proposed model was trained to predict longitudinal velocity, lateral velocity, and yaw rate based on the current vehicle state and control inputs. In addition, multi-step self-feedback training was applied to reduce accumulated errors during closed-loop rollout, and its performance was compared with that of a single-step prediction model. The experimental results show that the multi-step training model had a slower increase in error during long-term prediction. In particular, it showed more stable performance in predicting longitudinal and lateral velocities. The analysis of prediction window length confirmed that using an appropriate window length improved the balance between training stability and long-term prediction accuracy.
Authors
- Sanghyeon Park
- Chanho Park
- Dae-Hyun Kim
Institutions
- Korea Railroad Research Institute (KR)
- Sungkyunkwan University (KR)
Publication Details
- Journal
- Journal of the Korean society for railway
- Published
- 2026-11-17
- DOI
- https://doi.org/10.7782/jksr.2026.29.11.1186
- Primary Topic
- Vehicle Dynamics and Control Systems
- Type
- article
- Field-Weighted Citation Impact
- 0.00