// the find
louiszengCN/CarlaAir
CarlaAir: Fly Drones Inside a CARLA World!! A Unified Infrastructure for Air-Ground Embodied Intelligence
CARLA-Air fuses CARLA's urban driving simulation with AirSim's multirotor flight inside a single Unreal Engine 4 process, eliminating the inter-process bridge that makes co-simulation painful. It targets robotics researchers who need air-ground scenarios — drone surveillance of traffic, VLN data collection, RL training — without cobbling together two separate simulators. Only ~35 lines of upstream code were changed to make both APIs coexist.
The single-process architecture is the real differentiator: both the CARLA Python client and AirSim client connect to the same world, so weather changes, actor spawns, and sensor timestamps are all naturally synchronized with no bridge latency. The coordinate alignment between CARLA's left-handed system and AirSim's NED frame is documented and verified to sub-millimeter precision, which is the kind of thing that quietly breaks research if you get it wrong. Pre-built binaries for Ubuntu 20.04/22.04 and Windows 11 with a one-script env setup lower the barrier significantly — you can actually run a drone chase demo in four commands. The 89/89 CARLA API test pass rate and 3-hour/357-reset stability numbers are concrete, not just aspirational.
It's locked to UE4.26 and CARLA 0.9.16, which is already aging; CARLA itself has been moving toward UE5 and this fork will fall further behind. Building from source is not for the faint-hearted — the CARLA build process is notoriously painful and nothing here changes that, so you're dependent on the maintainers keeping pre-built binaries current. The non-commercial license on the badge contradicts the MIT claim in the license section, which is going to cause confusion for anyone who needs to use this in a funded research context. Multi-drone support is still on the roadmap, meaning the 'air-ground' story is currently one drone plus traffic, not swarms or heterogeneous fleets.