Vehicle Development and Testing - Data-driven Method to Reduce Error of the Simulation’s Control Response of the Real Vehicle

With the steady increase of autonomous vehicles numbers and advanced driver-assistance system features aimed at improving road safety, simulation tools have become a critical part of the development process, enabling systems to be tested without risking property damage or bodily harm in the event of failure. Physics-based simulators are central to virtual vehicle development, yet their control responses often differ from real vehicles, limiting the transfer of controllers developed in simulation. As these simulations play a larger role in the vehicle design and validation process, a critical question is how well their predicted behavior translates to real-world physical systems becomes a factor.  This paper presents a calibration framework for an autonomous car platform that identifies and mitigates the residual error between simulated and real control responses. First, the error is quantified by issuing identical commands to a real vehicle and the simulated vehicle by monitoring actuator commands, wheel speeds, odometry, and the achieved motion of the real vehicle. Using this information, a data-driven model is trained that predicts the discrepancy between the simulator’s predicted state and the measured real response. The learned error is injected back into Gazebo by a ROS 2 plugin that applies corrective effects at each step, preserving stability while compensating for unmodeled dynamics. The model is then assessed by repeating the original experiments to determine accuracy gains, reporting terminal position error, velocity, and acceleration RSME. The contribution is a practical, reproducible, control- oriented simulation of a real-world calibration pipeline for car dynamics to increase the control-response fidelity using the Gazebo simulator and the ROS 2 framework.