CARLA 0.9.11 Release (2025)

What better way to round off the year than with the release of CARLA 0.9.11?

In this release there has been a big focus on improving determinism, with the goal ofmaking CARLA more reliable and stable. Traffic Manager can now be used in full deterministic mode, and eventhe animations used in pedestrian collisions (rag dolls) are deterministic by default.

CARLA 0.9.11 brings many fixes and updates of critical features. The PhysX Vehicle Manager has beenenhanced to provide more realistic wheel rolling physics and accurate collisions.We have extended the maps API to provide the ability to toggle on and off every environment object individually. We also havenew versions of all our maps, in which users can load and unload the differentelements of the map by layers. The ROSbridge has plenty of improvements too, most notably the ability to spawn new pseudo sensors, whichallows the user to visualize and collect ground truth data.

There have been several new additions to the CARLA ecosystem too. We are proud to present thenew integration of CarSim (in beta). This is a powerful and high-fidelity vehicle dynamics simulatorthat allows users to perform realistic vehicle dynamics simulation of a specific vehicle model.
We are also excited to announce the integration of the Scenic traffic scenario specificationlanguage (in beta). And finally, we have added four new vehicle models to the blueprint library!

Here’s a rundown of the features that come with CARLA 0.9.11:

Improved PhysX Vehicle Manager - Sweep collisioncontrol improves the wheel rolling physics of our fleet of vehicles.
Map Sublevels - We created new optimized versions of our maps (tagged with the “Opt” suffix) that can be loaded and unloaded in a layer-by-layer fashion.
Toggle Environment Objects - Each object of the environment (asset) can be enabled and disabled in runtime.
Deterministic Traffic Manager - Traffic Manager is now fully deterministic.
Deterministic Rag Doll Animations - Pedestrian-vehicle collisions are deterministic by default.
New Vehicles - Four new vehicles added to the blueprint library.
Synchronous Mode Pipeline Update - An updated pipeline to fix delay in client server response times in synchronous mode.
CarSim Integration (Beta) - Support for co-simulation using CarSim (in beta), providing high-fidelity vehicle dynamics simulation.
Scenic Support (Beta) - Support for Scenic, a domain-specific traffic scenario language capable of generating challenging traffic situations in a few lines of code.
ROS Bridge Improvements - Spawn and destroy actors via ROSservices, generate pseudo sensors, new passive mode and integration of DVS camera, and semanticLIDAR.

Improved PhysX Vehicle Manager

The model controlling the interaction between vehicles’ wheels and surfaces has been upgraded,fixing an important problem that was causing wheels to intersect solid objects. Users will now have the optionto use a new model, based on sweep collision control. Sweep collision control allows forall possible contact planes along the wheel to be considered and thus ensure a smoother driveover obstacles and bumps in the scene. No more sudden jumps nor sinking into the curb!

Sweep collision control can be enabled using the following code:

physics_control = vehicle.get_physics_control()physics_control.use_sweep_wheel_collision = Truevehicle.apply_physics_control(physics_control)

Map Sublevels

New versions of all our maps have been released and boast some additional features. The new maps(identified with the suffix “Opt”) have the same layout as previous ones but they are now dividedinto sublevels. Thisallows users to loadand unload mapslayer by layer, providing additional control over the performance of the simulation. The available layers are as follows:

NONE - No layers selected.
Buildings - Sublevel containing all the buildings on the map.
Decals - Sublevel containing all the decals on the map.
Foliage - Sublevel containing all the foliage on the map.
ParkedVehicles - Sublevel containing all the parked vehicles on the map.
Particles - Sublevel containing all the particles on the map.
Props - Sublevel containing all the props on the map.
StreetLights - Sublevel containing all the streetlights on the map.
Walls - Sublevel containing all the walls on the map.
All - All layers selected.

Warning! Sublevels are only available in the "Opt" suffixed maps

Toggle Environment Objects

Users can now toggle individual environment objects in both old and new maps. Each objectin the environment has a unique ID that can be fetched and used to turn that object offand on.

Deterministic Traffic Manager

Users will now have the option to use Traffic Manager in deterministic mode. This will allowTraffic Manager to have the exact same output over distinct executions of a script as long asthe same conditions are maintained. Deterministic mode is easily turned on by providing an arbitrary seed value to the followingmethod:

my_tm.set_random_device_seed(seed_value)

The most important thing to keep in mind when using deterministic mode with Traffic Manager isthat both the world and Traffic Manager must be in synchronous mode to function properly. It is notpossible to achieve determinism in asynchronous mode due to the reduced amount of control thismode allows the client over the simulation. Read more about synchrony in CARLA and how to activate it for theworldand for Traffic Manager.

Deterministic Rag Doll Animations

Collisions with pedestrians trigger a set of animations to simulate damage. Traditionally,these animations have been directly controlled by Unreal Engine, which introduced a sourceof stochasticity. In this new version of CARLA we take direct control of the rag doll animations,making them deterministic by default. Deterministic rag dolls can be toggled viathe deterministic_ragdolls variable inthe carla.WorldSettings class. When set to true, pedestrians have less realistic death animationbut determinism is ensured. When false, simulation and collision are more realistic butdeterminism is not ensured.

New Vehicles

CARLA’s showroom now boasts 4 new vehicles! This brings the total number of available vehicles in theblueprint library up to 31. With these latest additions youcan now take a spin in our brand new Lincoln MKZ 2020, Mercedes-Benz Class C Coupe (by Fran Ruiz), DodgeCharger 2020 (by Martín Danlos) and Dodge Police Vehicle! These vehicles have been designed using new materials,with the goal of increasing the visual quality of the models.

Synchronous Mode Pipeline Update

We have developed an updated synchronisation pipeline that improves server-client communication and removes the frame delay present in previous versions.

Old synchronisation pipeline:

New synchronisation pipeline:

As shown in the above diagrams, we have changed the key moments in which synchronisation between the client and the server happens, providing a much quicker response from the client to the server. In the old pipeline, the client’s computations were based on the server’s information from the previous frame. This introduced a delay in the ability of the client to interact with the server. In the new pipeline, the input from the client is processed just before the physicscomputation. This, alongside the much earlier sensor rendering/computation, allows the client to respond in real-time and affect the server directly in the next frame.

CarSim Integration (Beta)

We are happy to announce the beta integration between CARLA and CarSim. CarSimis a simulation tool from Mechanical Simulation that performs highly accurate simulation of the dynamic behavior of passengervehicles and light-duty trucks. Using a 3D multi-body dynamics model, it reproduces the physicsof a vehicle in response to different controls (steering, throttle, braking, gear shifting) froma driver and/or automatic agent. This integration enables co-simulation between CarSim and CARLA, allowingusers to create CarSim vehicles, taking control over dozens of parameters, including suspension system, tires, and many more.

This first integration is great news for users of both CARLA and CarSim as it paves the way for themost powerful features of both platforms to be combined. Multiple vehicle simulations with collisioncapability on complex maps can be run in CARLA and then data can easily be analysed with CarSim’s highlyspecialised vehicle dynamics tools.

If you are interested in learning more about CarSim before using itwith CARLA, they have a great 2 part introduction to the tool on their YouTube channel.

Warning! This feature is still in experimental phase.

Scenic Support (Beta)

We are also proud to announce the integration of the Scenic programming language in CARLA. Scenic is a domain-specific probabilistic programminglanguage specifically created to easily define traffic scenarios. Our ongoing collaboration with the Scenic teamwill allow for straightforward definition and execution of traffic scenarios using CARLA. As part of the advantages of Scenic,we would like to highlight the capability of creating multiple and diverse instances of the same scenario from a single scenario definition.This will help users to focus on the generation of new traffic scenarios, reducing the amount of effort required to create variations ofan existing one.

Below, there is a code snippet of Traffic Scenario 03from the CARLA Autonomous Driving Leaderboard created with Scenic. Here we are using Scenic to create a situationwhere a pedestrian crosses the road unexpectedly and the ego vehicle is required to react safely.

# SET SCENARIO PARAMETERS AND MODELparam map = localPath('../../../tests/formats/opendrive/maps/CARLA/Town01.xodr')param carla_map = 'Town01'model scenic.simulators.carla.model # Here the definitions of all referenceables are defined (vehicle types, road library, etc)# SCENARIO CONSTANTSEGO_MODEL = "vehicle.lincoln.mkz2017"EGO_SPEED = 10SAFETY_DISTANCE = 10BRAKE_INTENSITY = 1.0PEDESTRIAN_MIN_SPEED = 1.0THRESHOLD = 18# BEHAVIORS:behavior EgoBehavior(speed=10): try: do FollowLaneBehavior(target_speed=speed) interrupt when withinDistanceToObjsInLane(self, SAFETY_DISTANCE): take SetBrakeAction(BRAKE_INTENSITY)behavior PedestrianBehavior(min_speed=1, threshold=10): do CrossingBehavior(ego, min_speed, threshold)## DEFINING SPATIAL RELATIONSlane = Uniform(*network.lanes)spot = OrientedPoint on lane.centerlinevending_spot = OrientedPoint following roadDirection from spot for -3# ACTOR CREATIONpedestrian = Pedestrian right of spot by 3, with heading 90 deg relative to spot.heading, with regionContainedIn None, # Allow the actor to spawn outside the driving lanes with behavior PedestrianBehavior(PEDESTRIAN_MIN_SPEED, THRESHOLD)vending_machine = VendingMachine right of vending_spot by 3, with heading -90 deg relative to vending_spot.heading, with regionContainedIn None # Allow the actor to spawn outside the driving lanesego = Car following roadDirection from spot for Range(-30, -20), with blueprint EGO_MODEL, with behavior EgoBehavior(EGO_SPEED)# REQUIREMENTSrequire (distance to intersection) > 30require (distance from vending_machine to intersection) > 30require always (ego.laneSection._slowerLane is None)require always (ego.laneSection._fasterLane is None)# TERMINATION CONDITIONterminate when (distance to spot) > 30

Execution of the script is easy and you can see some command line options in the Scenicdocumentation here.The following is an example of how we can run the above script:

scenic carlaChallenge3_dynamic.scenic --simulate

Watch this space for updates and improvements on our integration with Scenic and in the meantime,take a dive into the Scenic syntaxand explore all the possibilities!

Warning! This feature is still in experimental phase.

ROS Bridge Improvements

There have been quite a few improvements made to the ROS Bridge-CARLA experience. A new internalstructure has been developed to ease the integration of the ROS bridge with external tools. Actors can now be spawned and destroyed using ROS services, so we have added the SpawnObject and DestroyObjectservices. Now that actors can be spawned and destroyed in this manner, users candefine pseudo sensors. Pseudo sensors are much like normal sensors but instead of existing in the CARLAcontext they are developed directly in the ROS Bridge. This extension brings much more control over thedata collection process, as each user can define exactly what pseudo sensorsthey want to use.

The pseudo sensors available are:

sensor.pseudo.actor_list
sensor.pseudo.markers
sensor.pseudo.objects
sensor.pseudo.odom
sensor.pseudo.opendrive_map
sensor.pseudo.speedometer
sensor.pseudo.tf
sensor.pseudo.traffic_lights

Here is an example of how pseudo sensors can be defined:

{ "objects": [ { "type": "sensor.pseudo.opendrive_map", "id": "map" }, { "type": "vehicle.lincoln.mkz2017", "id": "ego_vehicle", "sensors": [ { "type": "sensor.camera.rgb", "id": "rgb_front", "spawn_point": {"x": 2.0, "y": 0.0, "z": 2.0, "roll": 0.0, "pitch": 0.0, "yaw": 0.0}, "image_size_x": 800, "image_size_y": 600 }, { "type": "sensor.pseudo.tf", "id": "tf" }, { "type": "sensor.pseudo.objects", "id": "objects" } ] } ]}

We have added a passive mode to the ROS Bridge to ease the integration of external toolswhen having multiple CARLA clients running at the same time. When enabling this mode, the ROSbridge stays passive and other clients have the responsibility of ticking and configuring the world.

Finally in the ROS improvements, the DVS Camera and the Semantic LIDAR are now fully integrated inthe ROS-Bridge.

Contributors

This space is dedicated to all of those whose contributions were merged in any of the project’sGitHub repositories during the development of CARLA 0.9.11. Thanks a lot for your hard work!

Changelog

Improved the documentation for use with pandoc tool by converting html tags to their markdown equivalent
Added parameter to carla settings to control culling
Refactored FAQ section of docs to use minimal html and fix broken layout
Added load_map_layer and unload_map_layer to control map layers on new maps that support subleveling
Added get_environment_objects call to get all the placed objects in the level
Added enable_environment_objects call to enable/disable objects of the level
Added fully deterministic option for Traffic Manager, sorting vehicles by ID and avoiding race conditions
Added the option to sweep the wheel shape for collision. This requires to patch the engine
Added the possibility of changing physics substepping options from client
Added ‘noise_seed’ to sensors to initialize the random generators
API extensions:
- Added actor.set_enable_gravity() function to enable/disable the gravity affecting the actor
Upgraded to DirectX 12 on Windows
Added horizontal_fov parameter to lidar sensor to allow for restriction of its field of view
Extended the local planner with a lateral offset.

Art

Added WorldSettings.deterministic_ragdolls to enable deterministic or physically based ragdolls

Fixes

Fixed RSSSensor python3 build and import of open drive maps by updating to ad-rss v4.2.0 and ad-map-access v2.3.0. Python import of dependent ‘ad’ python modules reflects now the namespaces of the C++ interface and follow doxygen documentation
Fixed sensor transformations and sensor data transformations mismatch in IMU and camera-based sensors
Fixed random dead-lock on synchronous mode at high frame rate
Fixed bug on Windows causing sun reflection artifacts
Fixed bug in waypoint.get_landmarks() causing some landmarks to be missed when s = 0
Fixed the actor.set_simulate_physics() for pedestrians and vehicles
Fixed bug causing camera-based sensors to stop sending data
Fixed the lack of determinism on the output of raycast sensors
Fixed bug in the actor’s id returned by the semantic lidar
Fixed error when using –config parameter in make package
Fixed dependency of library Xerces-c on package
Fixed minor typo in the simulation data section of the documentation
Fixed the config.py to read the .osm files in proper utf-8 encoding

← Previous Post Next Post →