AI Grand Prix — Autonomous Drone Racing Framework

Anduril AI-GP Competition | Virtual Qualifier Build

---

Architecture Overview

Vision + Telemetry → Perception → Planning → Control → MAVLink Commands → Simulator

This matches the exact control pipeline described in the official AI-GP Technical Specification (VADR-TS-001).

---

Project Structure

aigp/
├── main.py                  # Entry point — runs full autonomy stack
├── config/
│   └── settings.py          # All tunable parameters in one place
├── sim/
│   └── local_sim.py         # Local physics sim (no DCL connection needed yet)
├── perception/
│   └── gate_detector.py     # Gate detection from FPV camera frame
├── planning/
│   └── path_planner.py      # Next waypoint / gate sequencing
├── control/
│   └── drone_controller.py  # PID controller → TRPY commands
├── training/
│   └── video_trainer.py     # Extract gate frames from YouTube videos, build dataset
├── utils/
│   └── mavlink_bridge.py    # MAVLink v2 / MAVSDK UDP interface (for real sim)
└── logs/                    # Run telemetry logs

---

Modes

Mode	Command	Description
Local Sim	python main.py --mode sim	Physics sim + gate nav loop, no external connection
Video Training	python main.py --mode train --url <YouTube URL>	Extract gate frames from drone racing videos
MAVLink Live	python main.py --mode mavlink	Connect to DCL simulator via UDP (VQ1 launch)

---

Key Technical Spec Notes (VADR-TS-001)

• Interface: MAVLink v2 over UDP via MAVSDK-compatible interface
• Control messages: SET_POSITION_TARGET_LOCAL_NED or SET_ATTITUDE_TARGET
• Telemetry received: ATTITUDE, HIGHRES_IMU, ODOMETRY, HEARTBEAT, TIMESYNC
• Physics rate: 120 Hz | Command rate: 50–120 Hz recommended
• Vision: Forward-facing FPV camera (no depth, no GPS)
• VQ1: Desaturated env, highlighted gates, focus on completion
• VQ2: Complex lighting, no visual aids, fastest time wins
• Max run: 8 minutes

---

Quickstart

# Install dependencies
pip install -r requirements.txt

# Run local simulation
python main.py --mode sim

# Train perception on a drone racing YouTube video
python main.py --mode train --url "https://www.youtube.com/watch?v=EXAMPLE"

# Connect to DCL simulator (when VQ1 launches in May)
python main.py --mode mavlink --host 127.0.0.1 --port 14550

---

Competition Timeline

• VQ1: May 2026 — completion focus, <10 gates, simplified environment
• VQ2: June 2026 — speed focus, <20 gates, complex environment
• Physical Qualifier: September 2026, Southern California
• Grand Prix Final: November 2026, Ohio