We are thrilled to present to you the CARLA 0.9.16 release! This version brings some major upgrades to the Unreal Engine 4.26 version of CARLA that promise to augment your CARLA workflow and enhance the diversity of your simulation data!
CARLA 0.9.16 integrates powerful new rendering technologies and export options from NVIDIA. The new Cosmos Transfer1 foundational style transfer model allows the generation of multiple style variations of CARLA simulation output using text prompts, including variations in architecture, vehicles, weather and lighting conditions. Cosmos Transfer1 enables massive diversity augmentation of CARLA generated training datasets with only 1 set of simulation input data! NVIDIA NuRec brings neural reconstruction into the CARLA engine, allowing the re-rendering of a 3D scene learned from real-world sensor data with variations in viewpoint, camera configuration or perturbations in trajectory! NVIDIA’s SimReady Converter Tool enables the export of SimReady assets in the Universal Scene Description format (USD) for hassle-free transfer to other USD compatible applications from Omniverse.
Connectivity with ROS 2 is now natively supported by the CARLA server, abolishing the need for a bridge tool, giving a lower latency connection with ROS nodes for smoother simulations.
CARLA 0.9.16 includes support for Left-Handed Traffic! This long-awaited feature is finally here, allowing simulation of traffic for countries with left-handed traffic rules such as the UK or Japan. A new wheelchair model enables the inclusion of vulnerable road users into CARLA-generated training datasets.
Whether you’re building complex autonomy stacks, experimenting with digital twins, or exporting CARLA content to other platforms — this release brings you closer to production-grade simulation workflows.
🧪NVIDIA Cosmos Transfer1 integration
This CARLA release features NVIDIA Cosmos Transfer1 integration. Cosmos Transfer1 is a foundational style transfer model designed to augment simulation outputs.
With Transfer1, users can generate endless hyper-realistic video variations from CARLA sequences using simple text prompts. This capability is ideal for:
- Expanding visual diversity in perception datasets
- Bridging the domain gap for sim-to-real training
- Exploring edge cases with photorealistic textures, lighting, and weather variations
Furthermore, by using this feature in combination with Inverted AI’s DRIVE API, users can produce realistic behaviors enhanced with endless visual variations. The perfect combination for training AV stacks! You can read more about this topic here: NVIDIA: Acceleraging AV Simulation with Neural Reconstruction and World Foundational Models.
🎥 Neural Reconstruction with NVIDIA NuRec
This release also introduces support for NVIDIA NuRec 25.07, a state-of-the-art neural rendering pipeline.
With this integration, CARLA scenes can be rendered using a learned representation of light and geometry — enabling:
- Realistic rendering from arbitrary viewpoints
- Faster replays and view synthesis
- Neural graphics integration with synthetic datasets
This is a major step toward photorealistic simulation, especially for perception training and evaluation. Expect improvements over the next few releases as the neural rendering pipeline evolves.
Learn more about NuRec and how to use it in the CARLA documentation!
🧭 Native ROS2 Support
We’ve heard you — ROS2 is here.
CARLA 0.9.16 ships with native ROS2 integration, opening the door for:
- Out-of-the-box compatibility with modern robotic stacks
- DDS-based message passing and time synchronization
- Integration examples included
You can now connect CARLA directly to ROS2 Foxy, Galactic, Humble and more — with sensor streams and ego control - all without the latency of a bridge tool.
Autonomy teams, rejoice.
📦 SimReady USD Exporter
Need to export your CARLA environments or assets to other simulators or visualization platforms?
Our new USD SimReady Exporter lets you package your CARLA scenes and assets in the SimReady format — making them portable to the OpenUSD ecosystem and beyond.
Highlights:
- Export environment geometry, materials, textures, and metadata to USD
- Maintain physical properties and SimReady naming conventions
- Ideal for digital twin reuse, simulation federation, and visualization in tools like Omniverse or Isaac Sim
Whether you’re simulating, rendering, or training agents across multiple platforms, this tool reduces friction and increases reusability. Thanks to the NVIDIA team for this contribution!
Left-Handed Traffic Support
The CARLA 0.9.16 API includes support for Left-Handed Traffic. CARLA simulator now supports OpenDRIVE roads labelled with left-handed traffic rules, meaning that CARLA can accurately simulate traffic from left-handed driving countries such as the United Kingdom, India or Japan. We’re delighted to deliver this long-requested feature - we hope you find it useful!
Vulnerable Road Users
CARLA 0.9.16 includes a wheelchair model compatible with most of the existing pedestrian models, enabling the inclusion of vulnerable road users in CARLA generated training data. This helps avoiding the misclassification of wheelchair users as bicycle riders or users of other types of personal transport. Many thanks to Marc Teuber, Andreas Graf and Hamza Ben Haj Ammar from Itemis for this excellent community contribution!
# Choose a pedestrian blueprint
pedestrian_bp = blueprint_library.find('walker.pedestrian.0041')
# Check if the pedestrian model supports the wheelchair option
if pedestrian_bp.has_attribute('can_use_wheelchair'):
# Set the use_wheelchair attribute to true and spawn
pedestrian_bp.set_attribute('use_wheelchair','True')
pedestrian = world.spawn_actor(pedestrian_bp, spawn_point)
🗺️ Coming soon: Digital Twin Tool v2 - Build Your World From OpenStreetMap
We’ve completely re-engineered the Digital Twin Tool to make it easier than ever to generate full environments from real-world map data. Leveraging OpenStreetMap as the primary data source, this new version brings:
- Automatic parsing of road networks, intersections, and topology
- Heightmaps and semantic information from terrain datasets
- Flexible parameterization for traffic furniture, props, and buildings
- Templated urban layouts to fill in missing geometry while preserving georeference
With just a few clicks, users can generate a CARLA environment grounded in the real world — ideal for geo-specific experiments, regulatory testing, and sim-to-real pipelines. Now shipped as a standalone and easy-to-use tool.
This is a step toward bridging digital twins with procedural simulation. And it’s just the beginning.
🔧 Other improvements in 0.9.16
- Project Chrono update: local installation is now supported for Chrono version 6
- Python wheel support: Python versions 3.10, 3.11 and 3.12 are now supported
- Pythohn egg deprecation: Python eggs are now deprecated (only wheels are now provided)
- Masked materials in instance segmentation: Supports fine-grained annotations for complex objects (e.g. leaves, fences)
- Spline mesh fix - Resolved issue with invisible spline meshes in instance segmentation
- Inverted AI traffic example: Updated Python API script with support for waypoint-guided Inverted AI vehicles
- Debug drawing: Extended functions now allow rendering primitives directly on the HUD layer
- New vegetation props: 3 new tree models added to the props catalogue