skysim-demo1-preview-720p.mp4

Test drone autonomy in your browser — zero install.

Real aerodynamics in C++20. Real flight-controller firmware in the loop. ArduPilot · PX4 · Betaflight · Windows · Linux · macOS · WebAssembly

⬇ Download · ⬇ Download · 🧪 Benchmarks · 🗺 Roadmap · 🤝 Contribute

Every simulator that lets you test drone algorithms needs Linux, ROS, or a game engine and a GPU. SkySim's goal is to make it a URL.

Tested end-to-end: the Python agent interface, perception, data pipeline, benchmarks, and determinism harness; the Godot agent server, perception, and domain randomization (validated against Godot 4.3); and the web physics core (decoupled from Godot, validated natively).

Still needs the compiled C++ extension + hardware: real-flight aerodynamics, RGB rendering, cross-machine determinism numbers, and — the one thing no code can produce — a sim-to-real transfer result (see docs/sim_to_real.md). That last one is the single most valuable contribution open right now.

See ROADMAP.md for phase-by-phase progress.

0. Fly it in your browser — nothing to install

(Phase 4, preview)

👉 Open the live demo — works in any modern desktop browser (Chrome, Firefox, Edge, Safari).

The physics core compiles to WebAssembly and runs entirely client-side. Fly with the keyboard, or drop in a JavaScript control function. The demo is rebuilt and redeployed automatically on every push to main (.github/workflows/pages.yml). To host your own copy, run web/build_wasm.sh and serve web/. Heavy ML and firmware SITL use the native agent server; the browser tier is for interactive flight and classical/light control.

Grab a standalone build from the Releases page:

No Godot, no compiler, no dependencies. Everything is bundled.

Pick the zip for your OS from the Assets list on the latest release

The build is unsigned, so Gatekeeper blocks a normal double-click. Right-click (or Control-click) the app and choose Open, then confirm once.

Install Godot 4.3+ — one portable download, no admin rights, no installer needed.

From Releases, download SkySim-godot-project.zip and unzip it. Prebuilt extension binaries for every OS are already inside — you don't compile anything.

Open Godot → Import → select the unzipped folder → Play.

① Click Import ② Select the unzipped folder ③ Press Play (F5)

3. Developer / modder — build the C++ extension from source

You only need this if you're changing the C++ physics. It's one command and godot-cpp is fetched automatically — no git submodules.

Linux / macOS

./scripts/build.sh

Windows

scripts\build.bat

Prefer raw CMake? cmake -B build -DCMAKE_BUILD_TYPE=Release cmake --build build --parallel

Output: gdextension/bin/(lib)drone_sim.<platform>.release.<arch>.(dll|so|dylib)

Requirements: CMake ≥ 3.22, a C++20 compiler (MSVC 2022 / GCC 12+ / Clang 15+), Python 3.8+. Match godot-cpp to your editor with -DGODOTCPP_GIT_TAG=godot-4.3-stable.

Portability note: binaries are built without -march=native on purpose, so they run on any CPU of the same architecture. Turn on -DDRONE_SIM_NATIVE=ON only for a private single-machine build — it bakes in your CPU's instruction set and will crash elsewhere.

How the cross-platform binaries are made: you don't make them. Pushing a v* tag triggers GitHub Actions (.github/workflows/), which compiles the extension on Windows, Linux, and macOS (x86_64 + arm64), exports the standalone apps, and attaches everything to a Release automatically.

DroneBody (RigidBody3D subclass) ├── RotorArray → BladeElementSolver × N ← BET thrust/torque per rotor ├── Atmosphere ← ISA density/pressure + Dryden turbulence (MIL-HDBK-1797) ├── AeroEffectsBundle ← Cheeseman-Bennett ground effect + Leishman VRS ├── FlightController ← PX4-style cascade PID (attitude + rate) ├── MixerMatrix ← wrench → per-rotor throttle allocation ├── SensorSuite ← IMU / barometer / GPS with noise + bias └── SITLManager ← ArduPilot / PX4 / Betaflight bridges

Control priority each physics tick: SITL firmware (fresh actuator frame < 0.5 s) → set_rotor_throttles() → internal PID → motors off.

Copy the gdextension/ folder into your project root (Godot 4 auto-detects it — no plugin to enable).

Add a DroneBody node (appears under RigidBody3D) with a CollisionShape3D and a visual MeshInstance3D child.

For firmware-in-the-loop, add a SITLManager as a child of DroneBody.

Set Project Settings → Physics → Physics Ticks per Second = 250–400 (400 for ArduCopter). This project already ships at 400.

A minimal DroneBody scene tree Physics ticks set to 400

🔌 Flying with real flight-controller firmware (SITL)

Firmware Transport Port Who connects

ArduPilot UDP (sim = server) 9002 AP sends servo packets to us

PX4 TCP (sim = server) 4560 PX4 SITL connects to us

Betaflight TCP (sim = client) 5760 We connect to Betaflight

ArduPilot

cd ardupilot && sim_vehicle.py -v ArduCopter -f JSON:127.0.0.1 --console --map

PX4

cd PX4-Autopilot && make px4_sitl none_iris

Betaflight

cd betaflight && make TARGET=SITL && ./obj/main/betaflight_SITL.elf

Run the SkySim scene first, then start the firmware. Arming, modes, and missions come from your GCS (Mission Planner / QGroundControl). Frames are aerospace-standard (NED world / FRD body, see include/core/frames.hpp); rotor order follows the ArduPilot/PX4 quad-X convention.

SkySim on the left, QGroundControl on the right — the drone is flown by real PX4 firmware

tools/ap_json_probe.py and tools/px4_probe.py let you exercise the bridges without a full firmware checkout.

drone.get_telemetry() returns every physics tick: altitude, ground_speed, vertical_speed, roll/pitch/yaw_deg, roll/pitch/yaw_rate, total_thrust, power_draw, vrs_active, vrs_severity, ground_effect_factor, air_density, wind, sitl_active, sitl_source. Live PID tuning:

drone.set_rate_roll_pid(0.15, 0.05, 0.003) drone.set_rate_pitch_pid(0.15, 0.05, 0.003) drone.set_rate_yaw_pid(0.20, 0.10, 0.0)

🧪 Benchmarks & reproducibility

A small, stable suite of seeded tasks (hover, waypoint, gps_denied_nav) lets algorithms be compared and cited:

python examples/run_benchmark.py --port 5557 --out scorecard.json python examples/check_determinism.py --port 5557

For reference-quality reproducible numbers, build the extension with -DDRONE_SIM_DETERMINISTIC=ON (strict floating point). Details and how to submit results: docs/benchmarks.md.

Issues and PRs welcome — and you don't need to be a C++ physics person to help. A few places to start:

CI builds every PR on all three platforms, so if it compiles in the workflow it works for everyone. Physics stays in C++ (src/, include/); scene logic and dynamic-world helpers stay in GDScript (demo/).

If SkySim is useful to you, a ⭐ helps others find it.