AI Grand Prix Playground

An open-source, Elodin-based practice simulator for Anduril's AI Grand Prix, a $500K autonomous drone-racing competition. Built so contestants can iterate on perception, planning, and control code today, while the official Virtual Qualifier 1 simulator finishes baking.

What you get out of the box:

• High-fidelity 6-DOF physics from Elodin (deterministic, GPU-rendered, multi-rate sensors) around a generic 5-inch racing quad, our best public guess until the reference airframe is published.
• A real Betaflight SITL flight controller in lockstep with the physics, talking standard MAVLink-style RC + PWM over UDP.
• A forward FPV camera matching the AI Grand Prix tech-spec intrinsics (640×360, fx=fy=320, cx=320, cy=180, +20° up-tilt, 30 Hz).
• A 3-gate forward course in Elodin's ENU frame (+X/East), with automatic pass-time tracking.
• A clean solver/ package. That's the only directory you edit to compete.

Quick start (macOS / Ubuntu / Windows WSL, ~5 minutes)

You need uv, git, git lfs and a C toolchain for building Betaflight:

On macOS install the Xcode Command Line Tools (xcode-select --install).

On Windows:

• open Powershell as admin
• edit a new file with notepad "$env:USERPROFILE\.wslconfig"
• set the wsl network config with:

[wsl2]
networkingMode=mirrored

• and run wsl --install, which will start a new WSL distro

On Ubuntu/WSL run:

sudo apt update && sudo apt -y install build-essential clang-18 libasound2t64 git-lfs curl

Finally install uv:

curl -LsSf https://astral.sh/uv/install.sh | sh

Clone this repository and then:

# 1) Install the Elodin CLI (editor + run + db)
bash scripts/install_elodin.sh

# 2) Set up the Python environment
uv venv --python 3.13 && source .venv/bin/activate && uv sync

# 3) Fetch and build Betaflight SITL (one-time)
git submodule update --init --recursive --depth 1 betaflight
bash scripts/build_betaflight.sh

# 4) Configure Betaflight only if the committed eeprom.bin is missing or stale.
#    This is normally a no-op for a fresh clone.
uv run python scripts/configure_betaflight.py

# 5a) on Mac / Ubuntu, open the simulation in the Elodin editor to start sim and connect editor directly
elodin editor sim/main.py

# 5b) start just the simulation in WSL
elodin run sim/main.py
# then open a new Powershell, download & install the Elodin windows release binary
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/elodin-sys/elodin/releases/download/v0.17.3/elodin-installer.sh | sh
# clone the repo there as well, and run
elodin.exe editor 127.0.0.1:2240

After scripts/build_betaflight.sh, git status may show M betaflight. That is expected: the build script toggles Betaflight's target.h inside the submodule to enable simulator lockstep.

Expected output on a healthy run:

[SOLVER] using module: solver.baseline
[FPV] First frame at tick 33: 921600 bytes, shape=(360, 640, 4)
SUCCESS: SITL integration working! Drone took off!
[RACE] course=easy gates_passed=0/3 lap_time=15.00s status=DNF pass_times=[--,--,--]

Writing your solver

Edit solver/baseline.py or point at your own module:

RACE_SOLVER=my_team.my_solver elodin editor sim/main.py

The full contract (one autopilot(update: SensorUpdate) -> RCCommand function) is in solver/README.md.

Inspecting a run

Each run writes an auto-numbered betaflight_db### directory at the repo root (set ELODIN_DB_PATH to override). A couple of elodin-db commands you'll reach for after a run:

# Dump every committed component to flat, joined CSVs (one row per tick).
elodin-db export betaflight_db000 --format csv --flatten --join -o dbs/betaflight_db000-csv

# Render the FPV camera stream to a video file for offline inspection.
elodin-db export-videos betaflight_db000 -o betaflight_db000-video

The CSV form is convenient for regression diffs, offline plotting, or pulling state into a notebook. For lighter-weight queries on individual components without writing files, elodin-db query --eql ... hits the same DB; see ARCHITECTURE.md for the syntax.

Running tests

just test     # or: uv run pytest
just verify   # check the latest run log for a successful takeoff
just clean    # remove betaflight_db### directories and CSV/video exports

36 tests across packet round-trips, course geometry, camera intrinsics, and the baseline solver. All run in well under a second (no Elodin runtime needed).

Going deeper

• ARCHITECTURE.md: full design covering the lockstep cycle, coordinate frames, module reference, editor schematic, configuration knobs, and known opportunities for improvement.
• solver/README.md: the autopilot contract.
• The Elodin engine, editor, and Python bindings live at elodin-sys/elodin; read its README for the full surface area, including the sensor_camera and StepContext APIs we lean on here.

Acknowledgements

Built on top of the upstream elodin-sys/elodin/examples/betaflight-sitl, examples/sensor-camera, and examples/crazyflie-edu examples. Not affiliated with Anduril, DCL, Neros, or JobsOhio. The AI Grand Prix is theirs; this practice rig is community fan-art so contestants can start hacking now.

License

Apache 2.0, the same license as the upstream Elodin engine.