Configuring the simulation

Before you start running your own experiments there are few details to take into account at the time of configuring your simulation. In this document we cover the most important ones.

Changing the map

The map can be changed from the Python API with

world = client.load_world('Town01')

this creates an empty world with default settings. The list of currently available maps can be retrieved with


To reload the world using the current active map, use

world = client.reload_world()

Graphics Quality

Vulkan vs OpenGL

Vulkan (if installed) is the default graphics API used by Unreal Engine and CARLA on Linux.
It consumes more memory but performs faster.
On the other hand, OpenGL is less memory consuming but performs slower than Vulkan.


Vulkan is an experimental build so it may have some bugs when running the simulator.

OpenGL API can be selected with the flag -opengl.

> ./ -opengl

Quality levels

Currently, there are two levels of quality, Low and Epic (default). The image below shows how the simulator has to be started with the appropiate flag in order to set a quality level and the difference between qualities.

./ -quality-level=Epic ./ -quality-level=Low

Low mode runs significantly faster, ideal for users that don't rely on quality precision.

Running off-screen

In Linux, you can force the simulator to run off-screen by setting the environment variable DISPLAY to empty


DISPLAY= only works with OpenGL
Unreal Engine currently crashes when Vulkan is used when running off-screen. Therefore the -opengl flag must be added to force the engine to use OpenGL instead. We hope that this issue is addressed by Epic in the near future.

# Linux
DISPLAY= ./ -opengl

This launches the simulator without simulator window, of course you can still connect to it normally and run the example scripts. Note that with this method, in multi-GPU environments, it's not possible to select the GPU that the simulator will use for rendering. To do so, follow the instruction in Running without display and selecting GPUs.

No-rendering mode

It is possible to completely disable rendering in the simulator by enabling no-rendering mode in the world settings. This way is possible to simulate traffic and road behaviours at very high frequencies without the rendering overhead. Note that in this mode, cameras and other GPU-based sensors return empty data.

settings = world.get_settings()
settings.no_rendering_mode = True

Fixed time-step

The time-step is the simulation-time elapsed between two steps of the simulation. In video-games, this simulation-time is almost always adjusted to real time for better realism. This is achieved by having a variable time-step that adjusts the simulation to keep up with real-time. In simulations however, it is better to detach the simulation-time from real-time, and let the simulation run as fast as possible using a fixed time-step. Doing so, we are not only able to simulate longer periods in less time, but also gain repeatability by reducing the float-point arithmetic errors that a variable time-step introduces.

CARLA can be run in both modes.

Variable time-step

The simulation tries to keep up with real-time. To do so, the time-step is slightly adjusted each update. Simulations are not repeatable. By default, the simulator starts in this mode, but it can be re-enabled if changed with

settings = world.get_settings()
settings.fixed_delta_seconds = None

Fixed time-step

The simulation runs as fast as possible, simulating the same time increment on each step. To enable this mode set a fixed delta seconds in the world settings. For instance, to run the simulation at a fixed time-step of 0.05 seconds (20 FPS) apply the following settings

settings = world.get_settings()
settings.fixed_delta_seconds = 0.05


Do not decrease the frame-rate below 10 FPS.
Our settings are adjusted to clamp the physics engine to a minimum of 10 FPS. If the game tick falls below this, the physics engine will still simulate 10 FPS. In that case, things dependent on the game's delta time are no longer in sync with the physics engine. Ref. #695

Synchronous mode


Always run the simulator at fixed time-step when using the synchronous mode. Otherwise the physics engine will try to recompute at once all the time spent waiting for the client, this usually results in inconsistent or not very realistic physics.

The client-simulator communication can be synchronized by using the synchronous mode. When the synchronous mode is enabled, the simulation is halted each update until a tick message is received.

This is very useful when dealing with slow client applications, as the simulator waits until the client is ready to continue. This mode can also be used to synchronize data among sensors by waiting until all the data is received. Note that data coming from GPU-based sensors (cameras) is usually generated with a delay of a couple of frames respect to data coming from CPU-based sensors.

The synchronous mode can be enabled at any time in the world settings.

# Example: Synchronizing a camera with synchronous mode.

settings = world.get_settings()
settings.synchronous_mode = True

camera = world.spawn_actor(blueprint, transform)
image_queue = queue.Queue()

while True:
    image = image_queue.get()

For a more complex scenario synchronizing data from several sensors, take a look at the example

Command-line options


Some of the command-line options are not available in Linux due to the "Shipping" build. Therefore, the use of script is needed to configure the simulation.

Some configuration examples:

> ./ --no-rendering      # Disable rendering
> ./ --map Town05        # Change map
> ./ --weather ClearNoon # Change weather

To check all the available configurations, run the following command:

> ./ --help

Commands directly available:

  • -carla-rpc-port=N Listen for client connections at port N, streaming port is set to N+1 by default.
  • -carla-streaming-port=N Specify the port for sensor data streaming, use 0 to get a random unused port.
  • -quality-level={Low,Epic} Change graphics quality level.
  • Full list of UE4 command-line arguments (note that many of these won't work in the release version).


> ./ -carla-rpc-port=3000