Sim-to-Real Validation of a CARLA-Based Digital Twin for a 1:10 Scale Autonomous Vehicle

Abstract

Sim-to-real transfer remains a key challenge for the effective deployment of autonomous agents, as mismatches in perception, vehicle dynamics, and actuation can significantly degrade performance. We propose a low-cost, fully reproducible pipeline where a 1:10-scale Traxxas Slash vehicle serves as the physical twin of a model in the CARLA simulator, both operating under the same continuous action space. On the real platform, an NVIDIA Jetson Orin Nano executes the control policy and interfaces with an Arduino Due for sensor acquisition and PWM-based actuation. The digital twin is validated on a parking maneuver by executing the same reference policy in both domains and comparing yaw, lateral acceleration, and lateral offset using Pearson Correlation Coefficient (PCC) and Maximum Normalized Cross-Correlation (MNCC). Results confirm high-fidelity reproduction of the real platform dynamics. Building on this validation, a PPO reinforcement learning agent is trained entirely in simulation and deployed zero-shot on the physical vehicle. The policy achieves comparable performance, with only minor degradation primarily due to ultrasonic perception noise rather than dynamic mismatch. To the best of our knowledge, this is the first study demonstrating effective zero-shot transfer of a purely-simulation-trained autonomous scale vehicle.