Simulators are essential for training autonomous vehicle software, allowing developers to simulate millions of kilometers of driving scenarios without physical testing. While synthetic data is effective for testing trajectory planning and vehicle dynamics, it currently struggles to match real-world sensor noise and photorealism required for perception neural networks.
The history of traffic simulation began in the 1950s with models by Harry Goode and Martin Stark, which focused on intersection flow. By the 1970s, the industry adopted microscopic and network-level models like Japan's MISTRAN. Modern tools like SUMO are widely used for urban planning, though they lack the 3D physical detail necessary for autonomous driving systems.
The development of autonomous technology was further accelerated by DARPA challenges in the 2000s. Specialized robotics simulators such as Gazebo, Webots, and CoppeliaSim later emerged, providing realistic physics for sensor and control algorithm testing. These platforms integrated with the ROS ecosystem to bridge the gap between academic research and practical implementation.
The gaming industry significantly influenced the field by introducing reactive traffic in titles like Midtown Madness and GTA. This evolution culminated in 2017 with the release of CARLA, an open-source simulator built on Unreal Engine 4. Other notable tools include Microsoft's AirSim and NVIDIA's Isaac Sim, which leverage GPU acceleration for high-fidelity photorealistic environments.
Autonomous vehicle testing has evolved from basic object detection to complex multi-agent interaction and end-to-end driving challenges, as seen in the Waymo Open Dataset competitions from 2020 to 2025. Current simulation classes range from macroscopic traffic flow models to data-driven multi-agent systems like BehaviorGPT and VBD, which use real-world data to predict agent trajectories.
