CARLA — Open-Source Simulator for Autonomous Driving Research

Introduction

CARLA (Car Learning to Act) is an open-source simulator purpose-built for autonomous driving research. It provides photorealistic urban environments rendered by Unreal Engine, a flexible sensor suite, and a Python API for programmatic control. Researchers use CARLA to train and evaluate perception, planning, and control algorithms without deploying real vehicles.

What CARLA Does

• Renders photorealistic driving scenarios with dynamic weather, lighting, and traffic
• Simulates a comprehensive sensor suite including cameras, LiDAR, radar, GNSS, and IMU
• Provides a Python API for spawning vehicles, setting waypoints, and collecting sensor data
• Supports co-simulation with SUMO for realistic traffic flow modeling
• Includes a scenario runner for reproducible evaluation of driving agents against benchmarks

Architecture Overview

CARLA follows a client-server architecture. The server runs an Unreal Engine instance that renders the world and simulates physics. Clients connect via a TCP-based protocol using the Python or C++ API to control actors (vehicles, pedestrians, sensors) and retrieve simulation data. The server supports synchronous and asynchronous stepping modes, with synchronous mode guaranteeing deterministic replay.

Self-Hosting & Configuration

• Download prebuilt packages for Linux or Windows from the official releases
• Requires a GPU with Vulkan or DirectX support for rendering
• The Python client library installs via pip install carla
• Configure world settings (weather, map, traffic density) through the Python API
• Build from source with Unreal Engine 4 or 5 for custom map and asset development

Key Features

• Multiple detailed urban maps with intersections, highways, and residential areas
• Programmable weather and time-of-day for testing perception robustness
• Ground-truth labels for semantic segmentation, depth, and optical flow
• ROS bridge for integrating CARLA with Robot Operating System workflows
• Leaderboard challenge platform for standardized agent benchmarking

Comparison with Similar Tools

• LGSVL Simulator — discontinued in 2022; CARLA remains actively maintained with a larger community
• AirSim — focuses on drones and indoor environments; CARLA specializes in urban driving scenarios
• NVIDIA DRIVE Sim — commercial, high-fidelity driving simulator; CARLA is free and open source
• Waymax — lightweight, data-driven simulator by Waymo; CARLA provides full 3D rendering and physics

FAQ

Q: Can I train end-to-end driving models in CARLA? A: Yes. CARLA provides camera, LiDAR, and other sensor streams alongside vehicle controls, making it suitable for imitation learning and reinforcement learning pipelines.

Q: Does CARLA support ROS integration? A: Yes. The carla-ros-bridge package provides ROS topics for all CARLA sensors and controls, enabling integration with the ROS navigation and perception stacks.

Q: What maps are included? A: CARLA ships with several maps including Town01 through Town15, ranging from simple intersections to complex multi-lane highway interchanges.

Q: Can I create custom environments? A: Yes. You can build custom maps and assets using Unreal Engine and import them into CARLA. Documentation covers the full asset pipeline.

Sources

• https://github.com/carla-simulator/carla
• https://carla.org/