louiszengCN/CarlaAir

What it solves

CARLA-Air provides a unified simulation infrastructure for air-ground embodied intelligence. It solves the problem of simulating heterogeneous agents (like drones and autonomous cars) in a shared, high-fidelity urban environment without the latency and synchronization issues typically found in multi-process bridge co-simulations.

How it works

The project integrates CARLA (for urban driving and traffic) and AirSim (for physics-accurate multirotor flight) into a single Unreal Engine process. By modifying a small number of upstream files, it achieves absolute coordinate alignment between the two systems and allows a single Python script to control both APIs simultaneously. This architecture ensures that environmental changes, such as weather, affect both aerial and ground sensors in real-time with minimal communication overhead.

Who it’s for

It is designed for researchers and developers working on air-ground cooperation, embodied navigation (VLN/VLA), multi-modal perception dataset collection, and reinforcement learning (RL) for joint air-ground interactions.

Highlights

• Single-Process Composition: Combines CARLA and AirSim in one process to eliminate bridge latency.
• Synchronized Sensors: Supports up to 18 sensor modalities (including RGB, LiDAR, and Semantic Segmentation) aligned across both air and ground agents.
• Zero-Modification Migration: Existing CARLA and AirSim Python scripts and ROS 2 nodes can run without code changes.
• Realistic Urban Environment: Includes 13 urban maps with rule-compliant traffic and socially-aware pedestrians.
• Built-in Controls: Features FPS-style drone control (WASD + Mouse) directly in the viewport.
• ROS 2 Integration: Provides sensor bridges and RViz 2 presets for Humble-ready workflows.