carla
Open-source simulator for autonomous driving research.
CARLA is an open-source simulator for autonomous vehicle research that lets you test self-driving software in a virtual city with configurable sensors, traffic, and weather instead of real roads.
CARLA is an open-source simulator built for research into self-driving cars. Instead of testing driving software on real roads, which is slow, expensive, and risky, researchers can run it inside CARLA's virtual world of streets, buildings, and traffic. The project was created from the ground up to help people develop, train, and check autonomous driving systems in a safe, repeatable setting.
Along with the code, CARLA provides ready-made digital assets that you can use freely. These include city layouts, buildings, and vehicles that were made specifically for the simulator. The platform lets you decide what sensors a virtual car carries, for example cameras or other detectors, and lets you set the environmental conditions such as weather. That flexibility means a team can recreate the exact situations they want to study and repeat them as many times as needed.
CARLA is a demanding program to run. The recommended setup listed in the README includes a recent high-end Intel or AMD processor, at least 32 GB of memory, a powerful NVIDIA RTX graphics card, and Ubuntu 20.04. Ready-to-use builds are available for both Linux and Windows, and there are separate downloads for additional maps. People who want to build it themselves can clone the repository along with a special version of the Unreal Engine that CARLA depends on, then follow the build guides.
The project sits at the center of a wider set of related tools. These include a public leaderboard for validating driving software, a scenario runner for staging traffic situations, a bridge to the ROS robotics framework, and benchmarks for driving tasks. Detailed documentation, a Python programming interface, tutorials, and a frequently asked questions page are hosted online.
Licensing is split by component. CARLA's own code is released under the MIT license and its assets under a Creative Commons license, while some integrated dependencies and the Unreal Engine carry their own separate license terms that users should review. The README also asks that anyone using CARLA in published work cite the project's original 2017 research paper.
Where it fits
- Train and evaluate a self-driving algorithm in a simulated city without risk to real vehicles or people.
- Create repeatable traffic scenarios with specific weather and sensor configurations to benchmark autonomous driving software.
- Connect a ROS-based robotics pipeline to a virtual vehicle with a configurable camera and sensor rig.
- Run the CARLA public leaderboard evaluation to compare autonomous driving approaches using a standard benchmark.