AirSim is Microsoft Research's open-source drone and vehicle simulator, built on Unreal Engine 4. It shipped with a Python and C++ API, sensors covering most autonomous-systems research needs, and rendering good enough to train perception models on synthetic frames. Microsoft archived the public repo in 2022 and moved active development to a paid commercial successor called Project AirSim. The open code still compiles and runs. Whether that is enough for a new project in 2026 depends on what you are building.
Key Features
AirSim ships with the pieces you would expect from a research-grade simulator, plus a few that are unusual outside commercial products.
Unreal Engine 4 rendering
Scenes render at photoreal quality through UE4. For training computer vision models on synthetic data, this matters. A vision policy trained against a low-fidelity renderer will overfit to that renderer's artefacts and fail on real cameras. AirSim's frames are close enough to real footage that the domain gap becomes tractable rather than the dominant source of error.
Sensor coverage
RGB cameras, depth, segmentation, LIDAR, infrared, and event cameras are all first-class. Multi-camera rigs with configurable intrinsics are supported. If your research needs a sensor modality that a hobbyist sim skips, AirSim probably has it.
Python and C++ APIs
The Python API is the primary interface for most users. It exposes vehicle control, sensor reads, weather, time-of-day, and full state teleport. C++ works if you need lower latency or are integrating a compiled controller.
Flight controller and ROS integration
PX4, ArduPilot, and MavLink are all supported. ROS bridges exist. If the goal is software-in-the-loop testing for an actual airframe, this is closer to production than most alternatives.
Multi-vehicle and domain randomization
Multi-agent scenarios and randomized environments work out of the box. Both matter for reinforcement learning research that has to generalize past its training distribution.
Pricing Breakdown
AirSim is free under the MIT license. No feature gates, no seat limits, no telemetry.
| Plan | Price | What you get |
|---|---|---|
| Free (MIT) | $0 | Full source, all sensors, all vehicle types, Python and C++ APIs, Unreal Engine integration |
The real cost is hardware and time. UE4 wants a discrete GPU with 8 GB VRAM as a floor; 12 GB is more realistic once higher-quality environments are enabled. Initial setup, including Unreal Engine, Visual Studio build tools, and the AirSim plugin, takes several hours on a clean Windows machine and longer on Linux. Unreal Engine alone is a 100 GB install.
Pros and Cons
What works
- MIT license. Fork it, ship products on top, sell what you build. No lawyer conversation needed.
- Rendering quality good enough for perception training on synthetic data.
- Sensor list covers most autonomous-vehicle research needs.
- Multi-vehicle and domain randomization built in.
- PX4 and ArduPilot integration is close to what real airframes use.
What does not
- Archived by Microsoft in 2022. No security patches, no UE5 port, no ongoing bug fixes.
- Windows is the supported target. Linux builds compile but the plugin has known issues, and macOS is worse.
- Setup is a full-day task on a fresh machine.
- No hosted option. Every run consumes your local GPU.
- The commercial successor, Project AirSim, is a separate product with its own licensing and no upgrade path from the open version.
Who Is It For
AirSim makes sense in a few cases. A research group that already has an AirSim setup, published results against it, and needs reproducibility, since rebuilding on a different simulator invalidates prior comparisons. A project constrained to MIT-licensed dependencies for downstream distribution reasons, where a stricter license would be a blocker. A hobbyist with a capable Windows workstation and a spare weekend who wants to run drone RL experiments without paying for a cloud sim.
For everyone else, especially teams starting from zero in 2026, the archive status is a serious cost. Every dependency you build on that is no longer maintained becomes your maintenance problem eventually. UE4 itself is on wind-down; Epic released UE 5.5 in late 2024 and Unreal Engine 4 no longer receives substantive updates. AirSim's shelf life is bounded by how long the community keeps patching around that.
Verdict
AirSim scores 7.6/10. It is free and open source, with RPC APIs usable from several languages for images, state and vehicle control, plus PX4 and ArduPilot integration. The repo has 18,521 stars. It suits researchers with existing setups, and anyone reproducing earlier drone or car simulation work.
New projects should weigh the costs. Microsoft archived the project, 781 issues are open, and the licence shows as unclear (NOASSERTION). Setup needs Unreal Engine work, and no docs page was found at the checked address. For new autonomous driving research, CARLA is the better default per the earlier review. For pure drone work, PX4 software-in-the-loop with Gazebo is another route to check.