- SIM-CARLAAIR // W7
- Python
- ANIMA
SIM-CARLAAIR
One Simulator to Rule Air and Ground Together
Training and validating autonomous systems for joint air-ground operations requires two entirely separate simulation ecosystems: ground vehicle simulators (CARLA, Gazebo) and UAV simulators (AirSim, Flightmare). Bridging them for cross-domain scenarios — a convoy protected by drones, ground-guided UAV precision landing, multi-domain dataset collection — requires brittle inter-process glue code that breaks on every version update.
MODULE STATUS: PRODUCTIONWorkflow Coverage
5
- DIVISION
- GENERIC
- WAVE
- W7
- DOMAIN
- GENERAL
- WAVE 7 // ANIMA SUITE
- FOUNDATION — SIM-CARLAAIR
THE PROBLEM WE SOLVE
Training and validating autonomous systems for joint air-ground operations requires two entirely separate simulation ecosystems: ground vehicle simulators (CARLA, Gazebo) and UAV simulators (AirSim, Flightmare). Bridging them for cross-domain scenarios — a convoy protected by drones, ground-guided UAV precision landing, multi-domain dataset collection — requires brittle inter-process glue code that breaks on every version update.
Autonomous systems teams can iterate on joint air-ground AI policies in one coherent simulation environment — eliminating integration overhead and enabling the kind of multi-domain training scenarios that real-world testing cannot safely replicate.
WHAT SIM-CARLAAIR DELIVERS
SIM-CARLAAIR (codename: Ryujin) wraps CARLA-Air (arXiv:2603.28032), which fuses CARLA (ground vehicles, pedestrians, traffic management) and AirSim (drones, aerial sensors, multi-rotor physics) into a single shared Unreal Engine process with one unified Python API.
CAPABILITIES
- SIM-CARLAAIR (codename: Ryujin) wraps CARLA-Air (arXiv:2603.28032), which fuses CARLA (ground vehicles, pedestrians, traffic management) and AirSim (drones, aerial sensors, multi-rotor physics) into a single shared Unreal Engine process with one unified Python API
- Five ready-to-run workflows cover precision landing, VLN/VLA evaluation, large-scale dataset collection, cross-view perception benchmarking, and RL environment instantiation
- Three custom CUDA kernels deliver 18.6x speedup on coordinate transforms and 40.8x on multi-view compositing.
WHY THIS IS HARD
Building SIM-CARLAAIR requires solving multiple coupled problems:
- 01Training and validating autonomous systems for joint air-ground operations requires two entirely separate simulation ecosystems: ground vehicle simulators (CARLA, Gazebo) and UAV simulators (AirSim, Flightmare)
- 02Bridging them for cross-domain scenarios — a convoy protected by drones, ground-guided UAV precision landing, multi-domain dataset collection — requires brittle inter-process glue code that breaks on every version update.
- 03SIM-CARLAAIR (codename: Ryujin) wraps CARLA-Air (arXiv:2603.28032), which fuses CARLA (ground vehicles, pedestrians, traffic management) and AirSim (drones, aerial sensors, multi-rotor physics) into a single shared Unreal Engine process with one unified Python API
- 04Five ready-to-run workflows cover precision landing, VLN/VLA evaluation, large-scale dataset collection, cross-view perception benchmarking, and RL environment instantiation
SIM-CARLAAIR solves these through careful architecture design and rigorous validation.
PROOF, NOT PROMISES
Key performance metrics:
| METRIC | VALUE |
|---|---|
| Workflow Coverage | 5 ready-to-run mission profiles (precision landing, VLN/VLA, dataset, cross-view, RL) |
| CUDA Acceleration | Coordinate transform 18.6x faster; multi-view compositor 40.8x faster vs. CPU baseline |
| Integration Surface | 63-topic ROS2 bridge + FastAPI server — plugs into any autonomy stack |
| Defense Angle | Joint air-ground mission rehearsal, multi-domain AI training, counter-UAS sensor fusion development |
WHAT'S BUILT TODAY
| COMPONENT | STATUS | NOTES |
|---|---|---|
| Workflow Coverage | COMPLETE | 5 ready-to-run mission profiles (precision landing, VLN/VLA, dataset, cross-view, RL) |
| CUDA Acceleration | COMPLETE | Coordinate transform 18.6x faster; multi-view compositor 40.8x faster vs. CPU baseline |
| Integration Surface | IN PROGRESS | 63-topic ROS2 bridge + FastAPI server — plugs into any autonomy stack |
| Defense Angle | IN PROGRESS | Joint air-ground mission rehearsal, multi-domain AI training, counter-UAS sensor fusion development |
WHERE SIM-CARLAAIR DEPLOYS
- APP_01
AUTONOMOUS SYSTEMS
Autonomous systems teams can iterate on joint air-ground AI policies in one coherent simulation environment — eliminating integration overhead and enabling the kind of multi-domain training scenarios that real-world testing cannot safely replicate.
- APP_02
RESEARCH LABS
SIM-CARLAAIR (codename: Ryujin) wraps CARLA-Air (arXiv:2603.28032), which fuses CARLA (ground vehicles, pedestrians, traffic management) and AirSim (drones, aerial sensors, multi-rotor physics) into a single shared Unreal Engine process with one unified Python API.
- APP_03
EDGE COMPUTING
Merge drone and ground vehicle simulation in one Unreal Engine process — 18.6x faster coordinate transforms, 63-topic ROS2 bridge, zero config friction.