Data-Driven Calibration of Vehicle Simulations for Accurate Control Response in Development and Testing

With the steady increase in autonomous driving (AD) and advanced driver-assistance systems (ADAS) aimed at improving road safety and navigation efficiency, simulation tools have become a critical part of the development process, allowing systems to be tested while mitigating the risk of physical injury or property damage upon failure. Physics-based simulators are central to virtual vehicle development, yet their control responses often differ from real vehicles, potentially limiting the transfer of controllers and algorithms developed in simulation. As these simulations play an important role in the vehicle design and validation process, a critical question is how well their predicted behavior translates to real-world physical systems.

This paper presents a calibration framework for an autonomous vehicle platform that learns the motion characteristics of an experimental vehicle and uses that knowledge to correct the actuator response of a simulation model. The model is trained by collecting training data consisting of angular velocity data recorded during motion sequences from the wheel encoders mounted on the vehicle. After the data is post-processed, a long-short-term memory (LSTM) model is trained that predicts the angular velocity that the physical vehicle would achieve given a sequence of the past 30 command velocities.