The evolution of autonomous systems is moving beyond isolated domains, pushing researchers to look for environments where aerial drones and ground vehicles can interact seamlessly. CARLA-Air emerges as a critical bridge in this space, integrating high-fidelity urban driving with physics-accurate drone flight within a single, unified Unreal Engine framework. By merging these previously separate worlds, the platform eliminates the synchronization issues that often plague co-simulations where different software components attempt to communicate across boundaries.

What makes this release particularly significant is its commitment to spatial-temporal consistency, meaning every object in the virtual world moves and reacts at the exact same time-step (physics tick). This precision is vital for training embodied intelligence—AI agents that learn by interacting with their physical surroundings rather than just processing static data. Whether it is a drone scouting a path for a delivery robot or a fleet of vehicles navigating a busy city, the simulation provides 18 different sensor modalities, from LiDAR to cameras, to capture every detail of the environment.

Furthermore, CARLA-Air acts as a spiritual successor to AirSim, a widely used flight simulator that was recently archived. By preserving the native Python interfaces and ROS 2 support that researchers already rely on, the project allows for zero-modification code reuse. This lowers the barrier to entry for developers working on everything from reinforcement learning to vision-language models, ensuring that the legacy of open-source aerial robotics continues to thrive in a modern, scalable infrastructure.

The evolution of autonomous systems is moving beyond isolated domains, pushing researchers to look for environments where aerial drones and ground vehicles can interact seamlessly. CARLA-Air emerges as a critical bridge in this space, integrating high-fidelity urban driving with physics-accurate drone flight within a single, unified Unreal Engine framework. By merging these previously separate worlds, the platform eliminates the synchronization issues that often plague co-simulations where different software components attempt to communicate across boundaries.

What makes this release particularly significant is its commitment to spatial-temporal consistency, meaning every object in the virtual world moves and reacts at the exact same time-step (physics tick). This precision is vital for training embodied intelligence—AI agents that learn by interacting with their physical surroundings rather than just processing static data. Whether it is a drone scouting a path for a delivery robot or a fleet of vehicles navigating a busy city, the simulation provides 18 different sensor modalities, from LiDAR to cameras, to capture every detail of the environment.

Furthermore, CARLA-Air acts as a spiritual successor to AirSim, a widely used flight simulator that was recently archived. By preserving the native Python interfaces and ROS 2 support that researchers already rely on, the project allows for zero-modification code reuse. This lowers the barrier to entry for developers working on everything from reinforcement learning to vision-language models, ensuring that the legacy of open-source aerial robotics continues to thrive in a modern, scalable infrastructure.