CarlaAir (louiszengCN/CarlaAir: CarlaAir: Fly Drones Inside a CARLA World!! A Unified Infrastructure for Air-Ground Embodied Intelligence)
Objective: Have a simple, yet powerful simulator that can represent UAV and UGV in outdoor environments under different conditions and record data using multiple sensor modalities.
Simple but powerful
Able to run/start with minimal intermediate steps (don’t have to be an übernerd to run)
Note that SkyScenes also might be able to run inside of CarlaAir, as Carla API remains unchanged
UxV representation
It must be able to accurately represent UxVs
Outdoor scenes
It must be able to accurately represent outdoor environments in various conditions, including adverse ones, such as rain, night, etc.
Scripts
There should be a collection of clear, logically defined scripts that provide various functionality
Must be simple yet not simplistic, as well as modular and maintainable
Data capture
There must be scripts that allow for the automated, reproducible capturing of data from UAV and UGV perspectives
Must be able to capture different sensor modalities including depth maps, ray maps, rgb images, and LiDAR
Additional requirements
No datasets are stored in the simulator, as it is out of scope
No evaluations are performed inside of the simulator, as they are out of scope
CarlaAir highlights
Single process composition (aligns with simple requirement)
Absolute coordinate alignment (aligns with data capture)
18 Sensor modalities (aligns with data capture)
Zero-modification code migration (aligns with simple requirement)
~20 FPS joint workloads (aligns with simulation capabilities and UxV representation)
3-hour stability verified (aligns with simple)
Built-in FPS drone control (aligns with UxV representation)
Realistic urban traffic (aligns with UxV representation and outdoor scenes)
Extensible asset pipeline (aligns with UxV representation and outdoor scenes)