MVP: autonomous flight without GPS/markers. Drone takes off, flies 4 waypoints using vision-based localization, then returns home. Training happens in simulation and deploys to a real ArduPilot drone with a Jetson companion.

• Sim: Gazebo + ArduPilot SITL
• Localization: ORB-SLAM3 (RealSense D455 stereo/depth)
• Control: MAVLink setpoints
• Orchestration: ROS2
• Learning: PPO baseline (vision + state)

• docs/: architecture, plan, and hardware notes
• sim/: simulation configs and scripts
• ros2/: ROS2 nodes (SLAM bridge, state, control)
• training/: RL training code and configs
• src/: shared utilities and wrappers

1. Bring up ArduPilot SITL + Gazebo and confirm manual flight.
2. Integrate RealSense + ORB-SLAM3 in sim and publish local pose.
3. Implement MAVLink setpoint node and fly a square in sim.
4. Train PPO to follow 4 waypoints using local pose and camera.
5. Deploy the same stack on Jetson and validate on a real drone.

See docs/architecture.md and docs/plan.md for details.