Open-Source Simulators for Airside Autonomous Vehicle Development

Research Date: 2026-03-22 Scope: Evaluation of open-source simulation platforms for autonomous vehicle development in airport airside environments (taxiways, aprons, ramps, gate areas).

CARLA Simulator (UE5.5)
AWSIM (Autoware/Unity)
NVIDIA Isaac Sim 5.x
LGSVL / SVL Simulator (Discontinued)
AirSim / Project AirSim
Gazebo (ROS Testing)
Comparison Matrix
Recommendations for Airside AV Development

1. CARLA Simulator (UE5.5)

Current State and Maintenance

CARLA is the most widely-used open-source autonomous driving simulator. It is actively maintained with two parallel development tracks:

Attribute	Detail
Latest stable release	0.9.16 (September 16, 2025)
Engine	Unreal Engine 5.5 (ue5-dev branch)
Legacy branch	UE 4.26 (ue4-dev branch, still available)
GitHub stars	~13,700
Contributors	161
Total commits (ue5-dev)	6,683+
License	MIT
Supported platforms	Ubuntu 22.04+, Windows 11
Hardware requirements	Intel i7/i9 9th gen+ or AMD Ryzen 7/9, 32GB+ RAM, NVIDIA RTX 3070+ (16GB+ VRAM)

The project is actively maintained. The ue5-dev branch receives nightly builds via GitHub Actions. Neya Systems has collaborated with the CARLA team on the UE5 migration. A prior version, 0.10.0, was also referenced in some documentation as a UE5-based release, but 0.9.16 represents the latest numbered stable release with full feature documentation.

Key 0.9.16 features:

NVIDIA NuRec 25.07 neural reconstruction integration
NVIDIA Cosmos Transfer1 style transfer for dataset augmentation
Native ROS 2 connectivity (no bridge required) with DDS-based message passing
USD SimReady Exporter for Omniverse/Isaac Sim interop
Left-handed traffic support (UK, Japan, India)
Wheelchair vulnerable road user model
Python 3.10/3.11/3.12 wheel support
Project Chrono 6 physics integration

Creating Custom Airport Environments in CARLA

Map Creation Workflow

CARLA requires two files for any custom map:

.xodr -- OpenDRIVE road network definition
.fbx -- 3D geometry mesh with textures

Recommended tool: MathWorks RoadRunner (commercial, not free) generates both formats. An open-source alternative is TrueVision Designer.

Import methods (source build):

Automatic make import (recommended)
RoadRunner plugin
Manual Unreal Engine import

Import method (packaged build, Linux only): Docker-based import, but no UE Editor customization.

Airport-Specific Challenges

Road network (OpenDRIVE): OpenDRIVE is designed for automotive road networks (lanes, junctions, signals). Airport taxiways and aprons do not map cleanly to this standard. Workarounds:

Model taxiways as single-lane roads with very low speed limits
Model apron areas as large parking zones or open-area junctions
Mark aircraft stands as parking spots
Use the make import method with custom .xodr files that encode taxiway centerlines as road segments

Key limitation: CARLA's traffic manager and AI-controlled vehicles assume road-legal driving behavior (traffic lights, lane keeping, intersection rules). Airport-specific logic (hold-short lines, pushback procedures, FOD avoidance) requires custom scripting via the Python API.

Adding Aircraft and GSE Actors

CARLA supports adding custom vehicle blueprints:

Create 3D model (50,000-100,000 triangles) with materials split into bodywork, glass, lights, interior
Rig to CARLA's base vehicle skeleton (hierarchy must be preserved; bone positions can be adjusted)
Create collision mesh (SMC_<name>.fbx) and raycast sensor mesh (SM_sc_<name>.fbx)
Import into Content/Carla/Static/Vehicles/4Wheeled
Create Animation Blueprint derived from VehicleAnimInstance
Create wheel blueprints and configure physics
Register in VehicleFactory blueprint (Make, Model, Class)

N-wheeled vehicles are supported -- configure additional wheel blueprints with differential torque distribution. This is relevant for multi-axle GSE like baggage tractors.

Limitation: The vehicle blueprint system is designed for cars and motorcycles. Aircraft (static obstacles) would need to be imported as static mesh props rather than dynamic actors, unless you implement a custom BaseVehiclePawn subclass with aircraft-specific physics. GSE with unconventional wheel arrangements or steering (e.g., aircraft tugs with 360-degree steering) will require significant custom work.

CARLA Sensor Models

Ray-Cast LiDAR (`sensor.lidar.ray_cast`)

Parameter	Default	Notes
Channels	32	Configurable
Range	10.0 m	Configurable
Points per second	56,000	Configurable
Rotation frequency	10.0 Hz	Configurable
Upper FOV	10.0 deg
Lower FOV	-30.0 deg
Horizontal FOV	360 deg
Atmosphere attenuation rate	0.004	Intensity loss per meter
Dropoff rate	0.45	Random point dropout proportion
Noise std dev	0.0	Along raycast vector

Hybrid Solid-State LiDAR (`sensor.lidar.hss_lidar`)

Parameter	Default	Notes
Channels	128	Based on Hesai AT128
Range	10.0 m
Rotation frequency	10.0 Hz
Upper FOV	12.9 deg
Lower FOV	-12.5 deg
Horizontal FOV	120 deg	Symmetric around forward
Horizontal resolution	0.1 deg

Matching RoboSense RS-HELIOS-5515 Specs

RS-HELIOS-5515 real-world specifications:

Parameter	RS-HELIOS-5515
Channels	32
Wavelength	905 nm
Range	150 m
Vertical FOV	70 deg (-55 to +15 deg)
Horizontal FOV	360 deg
Accuracy	+/-2 cm (1-100m), +/-3 cm (0.1-1m and 100-150m)
Data rate (single return)	576,000 pts/s
Data rate (dual return)	1,152,000 pts/s
RPM	600 / 1200
Dimensions	100mm dia x 100mm H
Weight	~1 kg
Power	9-32V
IP rating	IP67
Temp range	-30 to 60 deg C
Return modes	Single and dual
Beam distribution	Dense in center, sparse at edges

CARLA configuration to approximate RS-HELIOS-5515:

python

lidar_bp = blueprint_library.find('sensor.lidar.ray_cast')
lidar_bp.set_attribute('channels', '32')
lidar_bp.set_attribute('range', '150.0')            # 150m range
lidar_bp.set_attribute('upper_fov', '15.0')          # +15 deg
lidar_bp.set_attribute('lower_fov', '-55.0')         # -55 deg (70 deg total)
lidar_bp.set_attribute('rotation_frequency', '10.0') # 600 RPM = 10 Hz
lidar_bp.set_attribute('points_per_second', '576000') # Single return rate
lidar_bp.set_attribute('atmosphere_attenuation_rate', '0.004')
lidar_bp.set_attribute('dropoff_general_rate', '0.0') # Disable for clean data
lidar_bp.set_attribute('noise_stddev', '0.02')        # ~2cm noise

Fidelity gap: CARLA's ray-cast LiDAR does not model the RS-HELIOS's non-uniform beam distribution (dense center, sparse edges). All channels are evenly distributed across the vertical FOV. CARLA also lacks material-dependent reflectivity simulation in the standard ray-cast LiDAR (AWSIM handles this better with 32 configurable parameters vs. CARLA's 13). Dual-return mode is not natively supported in CARLA.

Camera Sensors

Type	Key Parameters
RGB (`sensor.camera.rgb`)	800x600 default, 90 deg FOV, bloom/lens flare/ISO/gamma/f-stop/shutter speed, exposure histogram or manual, lens distortion coefficients
Depth	R+G256+B65536 encoding, 1000m max range
Semantic segmentation	29 semantic classes, CityScapes palette
Instance segmentation	Semantic tag + unique instance ID
Wide-angle	Perspective, stereographic, equidistant, equisolid, orthographic, Kannala-Brandt projection models
DVS (event camera)	Asynchronous events, configurable positive/negative thresholds

Radar (`sensor.other.radar`)

Parameter	Default
Horizontal FOV	30 deg
Vertical FOV	30 deg
Points per second	1,500
Range	100 m
Output	Polar coordinates + velocity

Other Sensors

IMU: Accelerometer, gyroscope, compass with configurable noise per axis
GNSS: Lat/lon/alt with configurable bias and noise
Collision detector: Normal impulse data on collision events
Lane invasion detector: OpenDRIVE-based lane crossing detection
Obstacle detector: Ray-cast based, 5m default range
V2X (CAM and custom): Cooperative awareness messages, configurable transmission power/sensitivity/path loss

CARLA ROS Bridge

As of CARLA 0.9.16: Native ROS 2 connectivity is built into the CARLA server, eliminating the need for a separate bridge process. This provides:

DDS-based message passing with lower latency than the previous bridge architecture
Time synchronization between CARLA simulation clock and ROS 2 clock
Compatible with ROS 2 Foxy, Galactic, Humble, and later distributions
Sensor data published as standard ROS 2 messages (sensor_msgs/PointCloud2, sensor_msgs/Image, etc.)
Ego vehicle control via standard Autoware or custom command topics

Legacy ROS bridge (for older CARLA versions): The carla-simulator/ros-bridge package provides bidirectional ROS 1/ROS 2 communication. Community forks exist for specific version combinations (e.g., willv678/carla-ros2-bridge for CARLA 0.10.0+).

Autoware integration: The autoware_carla_interface package in Autoware Universe provides direct Autoware-to-CARLA connectivity. Uses carla_sensor_kit with 6 cameras (360-degree coverage), LiDAR, IMU, and GNSS. Known limitation: CARLA's Lanelet2 maps lack proper traffic light components for Autoware, and coordinate systems differ from the pointcloud map.

CARLA Leaderboard and Benchmark Agents

Current version: Leaderboard 2.1 (launched March 2025). Scoring shifted from exponential to linear infraction penalties, preventing agents from gaming the system by stopping early.

Status: The 2025 CARLA Autonomous Driving Challenge is not running due to unforeseen circumstances, but the leaderboard infrastructure remains available for local evaluation.

Evaluation framework:

Routes span freeways, urban, residential, and rural environments
Weather variations: daylight, sunset, rain, fog, night
NHTSA typology scenarios: lane merge/change, intersections, roundabouts, traffic lights/signs, emergency vehicles, pedestrians, cyclists
Metrics: Driving Score (DS), Route Completion (RC), Infraction Penalty (IP)
Two modalities: SENSORS (perception-only input) and MAP (includes HD map access)
Hardware: AWS g5.12xlarge evaluation instances

Notable benchmark agents:

SimLingo -- 1st place on Leaderboard 2 SENSORS track
TransFuser++ / TransFuser v6 -- 2nd place at CVPR 2024 challenge; CVPR 2026 paper "LEAD: Minimizing Learner-Expert Asymmetry in End-to-End Driving"
InterFuser -- Zero collisions/km on leaderboard but 5.24 collisions/km under HABIT benchmark
BEVDriver -- Recent entrant using bird's-eye-view representations

CARLA for World Model Training Data

Neural reconstruction (NVIDIA NuRec): Converts recorded camera and LiDAR data into 3D scenes using neural networks. Enables re-rendering from arbitrary viewpoints, camera configurations, or trajectory perturbations. NuRec Fixer resolves reconstruction artifacts via a transformer-based inpainting model.

Cosmos Transfer1: Generates photorealistic video variations from CARLA sequences using text prompts. Can produce variations in architecture, vehicles, weather, and lighting from a single simulation run. NVIDIA's Physical AI Dataset release includes 40,000 Cosmos-generated clips.

Relevance to world models: CARLA + NuRec + Cosmos provides a pipeline for generating diverse, photorealistic training data. The workflow is: (1) run simulation scenarios in CARLA, (2) optionally apply Cosmos Transfer for style diversity, (3) use NuRec for neural 3D reconstruction, (4) render novel viewpoints and perturbations. This is directly applicable to training video prediction models, occupancy networks, and other world model architectures.

2. AWSIM (Autoware/Unity)

Overview

Attribute	Detail
Latest release	v2.0.1 (October 31, 2025)
Engine	Unity 6000.0.61f1
Rendering	HDRP or URP (switchable)
GitHub stars	~690 (TIER IV/AWSIM)
Commits	1,445+
License	Apache 2.0 (code), CC BY-NC (assets)
Forks	AWSIM-Labs (Autoware Foundation), D-AWSIM (distributed variant)
Maintenance	Active, latest release Oct 2025

AWSIM is purpose-built for Autoware development. It uses ROS2ForUnity for high-throughput, low-latency ROS 2 communication. A key fork, AWSIM-Labs, is maintained by the Autoware Foundation with additional features.

Sensor Simulation

LiDAR: GPU-accelerated via Robotec GPU Lidar (RGL) with ray tracing. 32 configurable parameters including material-dependent reflectivity -- significantly more detailed than CARLA's 13 parameters. This is a major advantage for matching real-world LiDAR behavior.
Camera: OpenCV-based with post-effects (bloom, motion blur, depth of field, noise, grain, color grading, glare, halo, sharpening, anti-aliasing). More post-processing options than CARLA.
IMU and GNSS: Standard simulation with configurable noise.
No radar support. This is a notable gap.

Custom Environment Creation

Requires three files:

Lanelet2 OSM file (lanelet2_map.osm) -- road network definition
PCD file (pointcloud_map.pcd) -- point cloud map for localization
3D mesh files (.fbx recommended, also .obj/.mtl/.png)

Import process:

Import .fbx into Unity Editor
Add Mesh Collider scripts for physics interaction
Import Lanelet2 and PCD files for Autoware compatibility
Add traffic rules and NPC behavior definitions

For airport environments: Lanelet2 is more flexible than OpenDRIVE for non-standard road geometries. Taxiways and apron areas can be represented as lanelets with custom regulatory elements. However, traffic behavior (NPC vehicles, pedestrians) still assumes road-driving conventions.

Vehicle Integration

Custom vehicle integration follows Autoware's standard workflow. Vehicle dynamics are optimized for Autoware's control systems. Logitech G29 steering wheel support is included for human-in-the-loop testing.

Performance

AWSIM maintains real-time simulation more reliably than CARLA. CARLA in asynchronous mode drops frames and is hard to control; in synchronous mode it runs below real-time with low ROS topic publishing rates. AWSIM supports configurable time scaling for flexible real-time factors.

Airport Capability Assessment

Feasible: Custom 3D airport environments can be imported. Lanelet2 is more suitable than OpenDRIVE for modeling taxiway networks. GPU-based LiDAR with material reflectivity is valuable for realistic apron surface detection.

Gaps: No radar sensor. NPC traffic assumes road vehicles. No built-in airport-specific assets. No aircraft or GSE models included.

3. NVIDIA Isaac Sim 5.x

Overview

Attribute	Detail
Latest release	v5.1.0 (October 21, 2025)
Platform	NVIDIA Omniverse
GitHub stars	~2,800
License	Open-source (Isaac Sim extensions); Omniverse Kit remains proprietary
Languages	Python (78.7%), C++ (18.2%)
Supported OS	Windows 10/11, Ubuntu 22.04 (x86_64 and aarch64)
Maintenance	Active

Key Features

Physics: GPU-accelerated PhysX, rigid body and vehicle dynamics, multi-joint articulation, SDF colliders
Rendering: RTX-based physically accurate rendering, neural reconstruction via NuRec and 3DGUT
Sensors: RTX and physics-based sensors defined via OmniSensor USD schema; stereo depth sensor with disparity artifacts
Synthetic data: Omniverse Replicator for large-scale SDG; Cosmos Transfer writer
ROS 2: Full ROS 2 Jazzy support, ZeroMQ bridge, MoveIt 2 tutorials, standardized ROS 2 simulation interfaces
Asset import: URDF, MJCF, CAD formats

MobilityGen

An extension for training data generation for mobile robots:

Generates occupancy maps, robot states, poses, velocities, images
Supports teleoperation, automated actions, customizable path planning
Works with differential drive robots, quadrupeds, humanoids
Custom robot registration via subclassing

Autonomous Vehicle Suitability

Isaac Sim is not designed for autonomous vehicles on public roads. Per NVIDIA's own forums, it lacks:

Realistic diverse outdoor scenarios (traffic conditions, weather, road types)
Complex vehicle dynamics modeling
Full AV sensor suites (LiDAR arrays, radar, multi-camera rigs)
Scalable large outdoor environments (urban/highway)

NVIDIA DRIVE Sim is the intended product for AV simulation but is not open-source and requires early access/enterprise licensing.

However, for airside AV (confined, structured environments): Isaac Sim is more viable than for on-road AV. Airport aprons are controlled, low-speed environments closer to warehouse/industrial settings -- which is Isaac Sim's sweet spot. Cyngn uses Isaac Sim for autonomous industrial vehicle deployment in warehouses with similar constraints.

Airport Capability Assessment

Strengths: USD-based environment authoring allows detailed airport scene construction. RTX sensor simulation is physically accurate. ROS 2 integration is production-ready. MobilityGen can generate training data for mobile robot navigation relevant to GSE operation.

Gaps: No built-in traffic simulation for multi-agent scenarios. No road network standards (OpenDRIVE/Lanelet2). No specific AV perception pipeline (no native BEV, no traffic light detection). Requires significant custom development for airport-specific scenarios.

Best use: Perception model training with high-fidelity synthetic data, especially for close-range obstacle detection and ground surface classification relevant to airside operations.

4. LGSVL / SVL Simulator (Discontinued)

Status

Attribute	Detail
Last active	January 2022
Developed by	LG Electronics Silicon Valley Lab
Engine	Unity HDRP
Status	Discontinued. No new releases, no PR reviews, no asset updates since early 2022
Web services	wise.svlsimulator.com shut down June 30, 2022
Source code	Still available on GitHub (`lgsvl/simulator`), buildable from source
License	Other (custom LG license)

What It Offered

High-fidelity Unity HDRP rendering
ROS 1/ROS 2 bridge
Autoware and Apollo integration
Sensor simulation (camera, LiDAR, radar)
Cloud-based scenario management
Custom map support

Why It Matters

LGSVL was the second-most-popular open-source AV simulator after CARLA. Its shutdown left a gap, particularly for Autoware users. The community response:

AWSIM was developed by TIER IV specifically to fill this gap
CARLA-Autoware-Bridge projects emerged to connect CARLA with Autoware
Some teams forked LGSVL source and maintain private builds

Alternatives

LGSVL Feature	Replacement
Autoware integration	AWSIM (primary), CARLA via autoware_carla_interface
Apollo integration	CARLA (direct support)
High-fidelity rendering	CARLA UE5.5, Isaac Sim RTX
Cloud scenario management	No direct open-source replacement; CARLA ScenarioRunner + OpenSCENARIO

5. AirSim / Project AirSim

History and Current Status

Attribute	AirSim (Original)	Project AirSim
Developer	Microsoft Research	IAMAI Simulations (ex-Microsoft engineers)
Status	Archived (2022)	Active development
Latest release	--	v0.1.1 (July 30, 2025)
Engine	UE4 / Unity	UE5
GitHub stars	~16,000 (microsoft/AirSim)	~563 (iamaisim/ProjectAirSim)
License	MIT	MIT
Platforms	Windows, Linux	Windows 11, Ubuntu 22
Focus	Drones + ground vehicles	Primarily drones/aerial autonomy

Architecture

Project AirSim has three layers:

Simulation libraries for generic robot structures
UE5 plugin connecting external physics and controllers
Client library for network-based API interactions

Custom physics, sensors, actuators, and controllers can be integrated. ROS support exists (/ros directory in repo).

Airport Relevance

Unique advantage: Project AirSim can create 3D environments from Bing Maps data, including a library of specific locations and generic spaces like airports. This is the only open-source simulator explicitly mentioning airport environment support.

Limitations:

Primarily designed for aerial vehicles (drones, air taxis)
Ground vehicle support exists but is not the primary focus
Small community (563 stars, 18 contributors)
Early-stage software (v0.1.1)
Sensor simulation details are sparse in documentation

Assessment for Airside AV

Project AirSim could be valuable for testing perception in airport environments (especially if using Bing Maps airport data), but it is not mature enough for full AV stack development. The aerial-first design means ground vehicle dynamics, traffic simulation, and sensor configurations are underdeveloped compared to CARLA or AWSIM.

6. Gazebo (ROS Testing)

Overview

Attribute	Detail
Latest LTS	Gazebo Jetty (Sep 2025 - Sep 2030)
Previous stable	Gazebo Harmonic, Gazebo Ionic
Engine	Custom (gz-sim), Ogre 2.x rendering, Vulkan support (Ionic+)
Physics	DART, ODE, Bullet (Featherstone API)
License	Apache 2.0
Maintenance	Active, maintained by Open Robotics

Gazebo Classic was officially discontinued and replaced by the modern "Gazebo" (formerly Ignition Gazebo).

Sensor Simulation

Camera (monocular, depth, thermal, segmentation, bounding box, wide-angle)
LiDAR (GPU ray-cast)
IMU, magnetometer, altimeter, airspeed
GPS/NavSat
Contact sensors, force-torque
Optical tactile sensor
Custom sensor plugins

ROS 2 Integration

ros_gz_bridge provides bidirectional message exchange between Gazebo Transport and ROS 2
Launch Gazebo from ROS 2 launch files
Integration with ros2_control for hardware abstraction
SDF robot description support
Model spawning and service bridging

Vehicle Simulation

Mecanum wheel controller
Tracked vehicle support
Wheel slip commands
Hydrodynamics for water vehicles
UAV support via PX4/ArduPilot SITL
Ackermann steering plugin available

Airport Capability Assessment

Strengths: Gazebo is the gold standard for ROS-based robot testing. SDF/URDF model format is well-suited for custom vehicles. Physics engines are accurate for low-speed ground vehicles. Lightweight enough to run many instances in CI/CD pipelines.

Gaps: Rendering quality is significantly below CARLA or AWSIM (Ogre 2.x vs. UE5 or Unity HDRP). No traffic simulation. No road network standards. No built-in AV perception pipeline. Not designed for photorealistic synthetic data generation.

Best use for airside AV: Unit testing of ROS 2 nodes, control algorithm validation, multi-robot coordination testing, and sensor integration verification. Not suitable for perception model training or end-to-end driving validation.

7. Comparison Matrix

Feature Comparison

Feature	CARLA 0.9.16	AWSIM v2.0.1	Isaac Sim 5.1	Project AirSim 0.1.1	Gazebo Jetty	LGSVL (archived)
Engine	UE 5.5	Unity 6000	Omniverse	UE 5	gz-sim/Ogre2	Unity HDRP
Maintenance	Active	Active	Active	Active (early)	Active	Discontinued
License	MIT	Apache 2.0	Open-source*	MIT	Apache 2.0	Custom
Latest Release	Sep 2025	Oct 2025	Oct 2025	Jul 2025	Sep 2025	Jan 2022

*Isaac Sim extensions are open-source; Omniverse Kit components are proprietary.

Rendering and Fidelity

Capability	CARLA	AWSIM	Isaac Sim	Project AirSim	Gazebo	LGSVL
Rendering quality	High (UE5 Lumen/Nanite)	High (HDRP/URP)	Very High (RTX)	High (UE5)	Medium (Ogre2)	High (HDRP)
Ray tracing	Yes	Yes (LiDAR)	Yes (full RTX)	Yes	No	Limited
Neural rendering	NuRec + Cosmos	No	NuRec + 3DGUT	No	No	No
Weather simulation	Rain/fog/night/cloudy	Limited	Configurable	Limited	Plugin-based	Yes
Dynamic lighting	Yes (Lumen)	Yes (HDRP)	Yes (RTX)	Yes	Basic	Yes

Sensor Simulation

Sensor	CARLA	AWSIM	Isaac Sim	Project AirSim	Gazebo	LGSVL
LiDAR	Ray-cast + HSS (13 params)	GPU ray-trace (32 params, material reflectivity)	RTX-based	Basic	GPU ray-cast	Ray-cast
Camera (RGB)	Yes (full post-processing)	Yes (extended post-processing)	Yes (RTX)	Yes	Yes	Yes
Depth camera	Yes	Yes	Yes (stereo disparity)	Yes	Yes	Yes
Semantic segmentation	Yes (29 classes)	Yes	Yes (Replicator)	Limited	Yes (plugin)	Yes
Radar	Yes	No	Limited	Limited	No	Yes
IMU	Yes	Yes	Yes	Yes	Yes	Yes
GNSS	Yes	Yes	Yes	Yes	Yes (NavSat)	Yes
Event camera (DVS)	Yes	No	No	No	No	No
V2X	Yes (CAM + custom)	Yes (V2I)	No	No	No	No

Simulation Speed and Performance

Metric	CARLA	AWSIM	Isaac Sim	Project AirSim	Gazebo
Real-time capability	Sync mode: below real-time; Async: above real-time with frame drops	Maintains real-time reliably	Depends on scene complexity	Real-time capable	Real-time for simple scenes
Time scaling	Fixed-delta configurable	Flexible time scaling	Configurable	Configurable	Configurable
Headless mode	Yes	Yes	Yes	Yes	Yes
Multi-instance	Possible but resource-heavy	Possible	Possible (cloud)	Limited	Yes (lightweight)
Min GPU	RTX 3070 (16GB VRAM)	RTX 2060+	RTX 3070+	RTX 2060+	Integrated OK
Disk footprint	~20 GB	~5 GB	~30 GB	~10 GB	<1 GB

Airport Airside Capability

Capability	CARLA	AWSIM	Isaac Sim	Project AirSim	Gazebo
Custom environment import	Yes (.fbx + .xodr)	Yes (.fbx + Lanelet2 + PCD)	Yes (USD, URDF, CAD)	Yes (Bing Maps, custom)	Yes (SDF, URDF)
Airport assets built-in	No	No	No	Airport env mentioned	No
Road network for taxiways	OpenDRIVE (poor fit)	Lanelet2 (better fit)	None (custom)	Bing Maps data	None (custom)
Custom vehicle types (GSE)	Yes (N-wheel, custom skeleton)	Yes (Autoware vehicle model)	Yes (URDF/USD)	Yes (custom)	Yes (SDF/URDF)
Aircraft as static obstacles	Yes (static mesh import)	Yes (Unity asset import)	Yes (USD import)	Yes (UE5 mesh)	Yes (SDF mesh)
Multi-agent traffic	Yes (Traffic Manager)	Yes (NPC traffic)	Limited	Limited	Limited (multi-robot)
Low-speed vehicle dynamics	Yes (PhysX / Chrono)	Yes (Unity physics)	Yes (PhysX)	Yes (custom physics)	Yes (DART/ODE/Bullet)

ROS Integration

Feature	CARLA	AWSIM	Isaac Sim	Project AirSim	Gazebo
ROS 2 support	Native (0.9.16+)	Native (ROS2ForUnity)	Yes (Jazzy)	Yes (/ros package)	Native (ros_gz_bridge)
ROS 1 support	Via legacy bridge	No	No	Via original AirSim	Via ros_gz_bridge
Autoware integration	autoware_carla_interface	Primary design target	No	No	Limited
Latency	Low (native DDS)	Low (ROS2ForUnity)	Low (ZeroMQ + DDS)	Medium	Low (native)
Standard message types	Yes	Yes	Yes	Partial	Yes

World Model Training Data Generation

Capability	CARLA	AWSIM	Isaac Sim	Project AirSim	Gazebo
Synthetic data pipeline	NuRec + Cosmos + standard sensors	Standard sensors, GPU LiDAR	Omniverse Replicator + MobilityGen + NuRec	Basic sensor recording	Basic sensor recording
Neural rendering	Yes (NuRec 25.07)	No	Yes (NuRec + 3DGUT)	No	No
Style transfer	Yes (Cosmos Transfer1)	No	Yes (Cosmos writer)	No	No
Domain randomization	Weather, lighting, textures	Limited	Extensive (Replicator)	Weather, lighting	Limited
Data format export	Custom + USD	ROS bags	USD + custom	Custom	ROS bags
Scalable generation	Yes (headless, multi-GPU)	Yes	Yes (cloud, multi-GPU)	Limited	Yes (lightweight)

Sim-to-Real Gap Assessment

Factor	CARLA	AWSIM	Isaac Sim	Project AirSim	Gazebo
Visual realism	High but imperfect in adverse weather	High	Very high (RTX)	High	Low
LiDAR realism	Moderate (no material reflectivity)	High (material reflectivity, 32 params)	High (RTX physics-based)	Low	Moderate
Physics accuracy	Moderate (PhysX/Chrono)	Moderate (Unity)	High (PhysX GPU)	Moderate	High (DART/Bullet)
Domain adaptation tools	Cosmos Transfer, NuRec	None built-in	Replicator domain randomization	None	None
Proven sim-to-real	Many published results (RALAD: 10-12% mAP improvement)	Limited published results	Published for manipulation, not AV	Published for drones	Published for manipulation/navigation
Key gap	Weather rendering, LiDAR material response	Radar absent, weather limited	Not designed for outdoor AV	Immature for ground vehicles	Visual fidelity too low

8. Recommendations for Airside AV Development

Primary Recommendation: CARLA 0.9.16

CARLA is the strongest choice for airside AV simulation due to:

Richest sensor suite including radar, V2X, and event cameras
Neural rendering pipeline (NuRec + Cosmos) for world model training data
Native ROS 2 without bridge overhead
Largest community (13.7k stars, 161 contributors, extensive published research)
Autoware integration via autoware_carla_interface
Leaderboard/benchmarks for measuring perception and planning performance
Custom vehicle support including N-wheeled configurations for GSE

Required custom work for airside:

Create airport environment in RoadRunner (or TrueVision Designer) with taxiway/apron geometry encoded as OpenDRIVE roads
Import aircraft 3D models as static props
Create GSE vehicle blueprints (baggage tractors, belt loaders, pushback tugs) with custom skeletons
Implement airport-specific traffic rules via Python API (hold-short logic, pushback sequences, right-of-way for aircraft)
Configure LiDAR to match RS-HELIOS-5515 specs (32 channels, 70 deg VFOV, 150m range, 576k pts/s)

Secondary Recommendation: AWSIM for Autoware Stack Testing

If the AV stack is built on Autoware, AWSIM provides:

Tightest Autoware integration (designed for it)
Superior LiDAR simulation (32 parameters, material reflectivity)
More reliable real-time performance
Lanelet2 map format is better suited for airport taxiway representation

Use AWSIM for closed-loop Autoware stack validation and CARLA for perception training data generation.

Supplementary Tools

Tool	Use Case
Isaac Sim	High-fidelity synthetic data generation for perception model training (close-range obstacle detection, surface classification). Use MobilityGen for navigation training data. Valuable if operating within NVIDIA's ecosystem (DRIVE Sim for later stages).
Gazebo	CI/CD testing of individual ROS 2 nodes, control algorithm unit tests, multi-robot coordination. Lightweight enough to run in automated pipelines. Not for perception training.
Project AirSim	Worth monitoring for airport environment generation from Bing Maps data. Too immature for production simulation today.

Sim-to-Real Strategy

The current state of the art for closing the sim-to-real gap in AV simulation (2025 research):

Cosmos Transfer (via CARLA 0.9.16): Generate photorealistic variations of simulated sequences to improve visual diversity
NuRec neural reconstruction: Convert real airport sensor recordings into interactive 3D scenes for replay with perturbations
Domain randomization: Vary lighting, weather, surface textures, and object placement systematically
RALAD framework: Retrieval-augmented learning bridges real-to-sim gap, improving mIOU by 10.3% and mAP by 12.3% while reducing retraining cost by 88%
Sim2Real Diffusion: Latent diffusion transforms simulated perception streams, bridging perceptual gap by 40%+

For airside specifically, the sim-to-real gap is potentially smaller than on-road AV because:

Airport environments are more controlled and structured
Speed is low (typically < 30 km/h)
Lighting conditions are more predictable (apron lighting)
Obstacle types are finite and well-defined (aircraft, GSE, personnel, FOD)
Surface textures are relatively uniform (concrete/asphalt)

The biggest remaining gap for airside will be LiDAR material response (jet blast deflectors, aircraft fuselage reflectivity, wet tarmac) -- where AWSIM's material-aware LiDAR or Isaac Sim's RTX sensors are superior to CARLA's standard ray-cast approach.

SLAM Methods

Methods

Open-Source Simulators for Airside Autonomous Vehicle Development ​

Table of Contents ​

1. CARLA Simulator (UE5.5) ​

Current State and Maintenance ​

Creating Custom Airport Environments in CARLA ​

Map Creation Workflow ​

Airport-Specific Challenges ​

Adding Aircraft and GSE Actors ​

CARLA Sensor Models ​

Ray-Cast LiDAR (sensor.lidar.ray_cast) ​

Hybrid Solid-State LiDAR (sensor.lidar.hss_lidar) ​

Matching RoboSense RS-HELIOS-5515 Specs ​

Camera Sensors ​

Radar (sensor.other.radar) ​

Other Sensors ​

CARLA ROS Bridge ​

CARLA Leaderboard and Benchmark Agents ​

CARLA for World Model Training Data ​

2. AWSIM (Autoware/Unity) ​

Overview ​

Sensor Simulation ​

Custom Environment Creation ​

Vehicle Integration ​

Performance ​

Airport Capability Assessment ​

3. NVIDIA Isaac Sim 5.x ​

Overview ​

Key Features ​

MobilityGen ​

Open-Source Simulators for Airside Autonomous Vehicle Development

Table of Contents

1. CARLA Simulator (UE5.5)

Current State and Maintenance

Creating Custom Airport Environments in CARLA

Map Creation Workflow

Airport-Specific Challenges

Adding Aircraft and GSE Actors

CARLA Sensor Models

Ray-Cast LiDAR (`sensor.lidar.ray_cast`)

Hybrid Solid-State LiDAR (`sensor.lidar.hss_lidar`)

Matching RoboSense RS-HELIOS-5515 Specs

Camera Sensors

Radar (`sensor.other.radar`)

Other Sensors

CARLA ROS Bridge

CARLA Leaderboard and Benchmark Agents

CARLA for World Model Training Data

2. AWSIM (Autoware/Unity)

Overview

Sensor Simulation

Custom Environment Creation

Vehicle Integration

Performance

Airport Capability Assessment

3. NVIDIA Isaac Sim 5.x

Overview

Key Features

MobilityGen