CARLA on ue5-dev: what the Unreal Engine 5.5 branch actually requires
Open-source simulator for autonomous driving research.
At a glance
- What is it?
- CARLA is an open source urban driving simulator for autonomous driving research, and the default branch ue5-dev is the Unreal Engine 5.5 line, not the UE 4.26 line. The build depends on a linked Epic Games account and a CARLA fork of Unreal Engine, which is the first thing to weigh before adopting it.
- Who is it for?
- Adopt CARLA if you need camera, lidar or radar data from a rendered urban environment and you can meet the stated floor of 32 GB RAM, 16 GB VRAM and Ubuntu 22.04 or 24.04 or Windows 11. Do not adopt it if your work is traffic flow modelling rather than perception and control, or if you cannot link a GitHub account to Epic Games, because the build stops there.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly C++, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The gap CARLA fills: rendered sensor data for driving stacks
Real driving data is expensive to collect and hard to reproduce. A logged run gives you one set of weather, one set of traffic, and no way to rewind and change a variable. CARLA takes the opposite position: it is a simulator built from the ground up, in the project's own words, to support development, training, and validation of autonomous driving systems. The output is not a traffic statistic. It is a rendered urban environment with configurable sensors and environmental conditions, so a perception model can be trained against the same corner with the sun moved, or a planner can be replayed against a different vehicle count.
The audience is narrow and specific. The topics list names deep learning, deep reinforcement learning, imitation learning and computer vision, and the ecosystem repositories back that up: a leaderboard for validating autonomous driving stacks, a scenario runner for executing traffic scenarios in CARLA 0.9.X, a ROS bridge, and separate repositories for conditional imitation learning and reinforcement learning. If your work is a control policy, a detector, or a scenario-based test harness, CARLA is aimed at you. If your work is signal timing or macroscopic congestion modelling, the rendered sensor layer is overhead you will pay for and never use.
The licence is MIT, and the README adds that the urban layouts, buildings and vehicles are open digital assets created for this purpose and usable freely. That combination matters for research groups that need to publish both code and the scenes behind a result.
How the pieces connect: server, Python client, and the asset layer
CARLA splits into a simulation server and client APIs. The server is a compiled Unreal Engine application that owns the world, the physics and the rendering. The client side is exposed through a Python API, documented separately from the C++ source, and the README points to a Python API reference and a blueprint library alongside the build guides. The blueprint library is the catalogue of prebuilt actors (vehicles, sensors and similar) that you spawn rather than construct from scratch.
Around that core sits a set of repositories that are not part of the main tree. Scenario_Runner executes traffic scenarios in CARLA 0.9.X. ROS-bridge connects CARLA 0.9.X to ROS. Driving-benchmarks provides benchmark tools for autonomous driving tasks. The leaderboard is described as an automatic platform to validate autonomous driving stacks. Two further bridges connect external stacks: carla-autoware for the AutoWare AV stack, and a reinforcement learning repository for running conditional RL models. Maps come in OpenDRIVE format, with RoadRunner named as a MATLAB GUI application for creating road networks and a separate standalone map editor for adding traffic light and traffic sign information to RoadRunner maps.
That layout tells you where to expect friction. Sensor specification is flexible, per the README, but the scenario and ROS layers are documented against 0.9.X, while the default branch is the UE 5.5 line. Version alignment between the core simulator and those satellite repositories is something you check yourself, not something the README resolves.
Building the ue5-dev branch: the Epic Games gate and the setup scripts
The README is explicit that ue5-dev is the development branch for the Unreal Engine 5.5 version of CARLA, existing in parallel with a UE 4.26 version on the ue4-dev branch, and it warns that there are significant differences between the two. The clone command names the branch directly:
git clone -b ue5-dev https://github.com/carla-simulator/carla.git CarlaUE5
Before that build can succeed you need access to the CARLA fork of Unreal Engine 5.5, which requires linking your GitHub account to Epic Games and then using your git credentials to authorise the download. This is the real gate. It is not a package you install from a registry.
On Linux, from the repository root, the interactive path is ./CarlaSetup.sh --interactive. The script prompts for your sudo password to install prerequisites, then prompts for GitHub credentials to authorise the Unreal Engine download. If you want to target an existing Python installation instead of the Python 3 the script installs via apt, pass --python-root=PATH_TO_PYTHON, using the output of whereis python3 with the /python3 suffix stripped.
For unattended Linux builds, the credentials go into an environment variable, exported in .bashrc as GIT_LOCAL_CREDENTIALS=username@github_token, and the script runs under sudo -E so the variable survives. The README also gives the inline form, sudo -E env GIT_LOCAL_CREDENTIALS=github_username@github_token ./CarlaSetup.sh. On Windows the entry point is CarlaSetup.bat, and unattended mode is stated as currently unavailable there, so you will be entering GitHub credentials or administrator privileges by hand. The README notes the Linux path may take some time and use a significant amount of disk space, without quantifying either.
Hardware and OS floors that rule machines out before you start
The recommended system section is unusually blunt, and it is the fastest way to decide whether CARLA is even an option. The CPU guidance covers Intel i7 9th to 11th generation, Intel i9 9th to 11th generation, AMD Ryzen 7 or Ryzen 9. Memory is listed as more than 32 GB. The GPU list is NVIDIA RTX 3070, 3080, 3090, 4090 or 5090, or better, with 16 GB or more of VRAM. Operating systems are Ubuntu 22.04 or 24.04, or Windows 11.
A second note removes any ambiguity: the Unreal Engine 5.5 version of CARLA will not work on Ubuntu 20.04 or Windows 10 or lower. This is a hard boundary, not a preference. A lab still running Ubuntu 20.04 LTS has to upgrade the host before the simulator enters the picture, and a workstation with 8 GB of VRAM is outside the stated range regardless of how small the scene is.
Read the list as a floor rather than a target. The README does not publish frame rates, sensor throughput or scene-size limits for any of these configurations, so treat the hardware table as the minimum the project is willing to name, not as a measure of what you will get. The VRAM figure in particular is the number to check against your actual sensor suite, because camera resolution and the number of active sensors are the variables that consume it.
Where CARLA is the wrong tool
The clearest limitation is environmental, and the README states it rather than burying it. If your infrastructure is pinned to Ubuntu 20.04 or Windows 10, the UE 5.5 branch is unavailable to you. The fallback is the ue4-dev branch, but the README frames that as a parallel version with significant differences, so it is a different target rather than a drop-in substitute. Code written against one branch is not guaranteed to behave the same on the other.
The second limitation is the build dependency itself. Access to the CARLA fork of Unreal Engine 5.5 runs through an Epic Games account link and git credentials. That is an organisational dependency as much as a technical one: an air-gapped machine, a shared build server without stored credentials, or a contributor without the account link cannot complete the setup. The unattended Linux mode exists precisely because that credential step is awkward, and Windows has no unattended mode at all.
The third is scope. CARLA renders sensors and simulates actors in an urban environment. It is not a traffic demand model, it does not produce macroscopic flow statistics, and the README makes no claim about calibration against real-world traffic volumes. If your research question is about network-level congestion rather than what a camera or lidar sees, the rendering stack is cost without benefit. The README also gives no figures on build duration, asset size or simulation throughput, so anyone who needs to plan capacity around those numbers has to measure them locally.
SUMO is the different tool, not the better one
The obvious comparison for anyone weighing CARLA is SUMO, and the difference is architectural rather than a matter of degree. SUMO models traffic: vehicles are entities in a road network with car-following and lane-changing behaviour, and the output is flow, travel time and emissions at network scale. There is no renderer producing camera or lidar frames, because the question it answers is about the movement of many vehicles, not about what one vehicle perceives.
CARLA inverts that. The README describes a platform supporting flexible specification of sensor suites and environmental conditions, with the Python API and blueprint library as the interface for spawning actors and reading sensors. The unit of interest is the ego vehicle and its sensor rig inside a rendered urban scene. The scene is the point, not the aggregate.
That said, the two are not mutually exclusive in practice, and the CARLA ecosystem shows the seams. Maps arrive in OpenDRIVE format, the same road network description SUMO consumes, and RoadRunner is named as the tool for creating those networks. So the road geometry can be shared even though the simulation layers cannot. If you need both aggregate traffic behaviour and rendered perception, you are looking at two simulators and a conversion step, not one tool that does both. The README does not describe a supported SUMO co-simulation path, so treat any such integration as work you own.
Maintenance, versions and the MIT licence
The release history in the repository shows an uneven cadence. 0.9.15 landed in November 2023, 0.10.0 in December 2024, and 0.9.16 in September 2025, with the ue5-dev branch receiving pushes as recently as September 2026. Two version lines are being maintained in parallel, which is the practical maintenance cost: fixes and features do not automatically appear on both, and the README's warning about significant differences between UE 5.5 and UE 4.26 means a project pinned to one branch cannot assume parity with the other.
Rebuilding is the other recurring cost. The README states that once setup is complete, subsequent builds can be executed with commands run from the CARLA root directory, which implies the Unreal Engine download and prerequisite installation are a one-time expense and later builds are cheaper. The initial setup is where the time and disk space go, and the README declines to estimate either.
The licence is MIT, which is permissive and imposes no copyleft obligation on your own code. That is the extent of what the supplied material supports. The README separately states that the urban layouts, buildings and vehicles are open digital assets created for this purpose and usable freely, which is a claim about the assets rather than a licence grant, and the two should not be conflated. If you plan to redistribute modified assets or ship a product built on the simulator, read the actual licence file and the asset terms in the repository rather than relying on a summary. Nothing here is legal advice.
Editorial conclusion
Adopt CARLA if you need camera, lidar or radar data from a rendered urban environment and you can meet the stated floor of 32 GB RAM, 16 GB VRAM and Ubuntu 22.04 or 24.04 or Windows 11. Do not adopt it if your work is traffic flow modelling rather than perception and control, or if you cannot link a GitHub account to Epic Games, because the build stops there. Before committing, verify three things: that the ue5-dev branch matches the API surface your existing code targets, that the disk space and build time for the Unreal Engine 5.5 download are acceptable on your machine, and that the assets you need appear in the published asset catalogue.
Community notes