No map, no GNSS, no compass, no lidar: autonomous drone flies a building on two cameras

Autonomous point-to-point flight through a complex 3D urban location (Urban Circuit Practice 01) with no static map, no GNSS, no magnetometer and no lidar. The drone knows only its start and goal; everything else comes from a forward stereo pair and two small time-of-flight sensors. The same pair does both jobs: it builds the obstacle memory the planner searches, and a stereo visual-inertial filter on those images is the position and heading the autopilot flies on. What you see: Gazebo: the vehicle, the environment, the follow camera. RViz: blue is the current stereo depth return, pink is the accumulated 3D obstacle memory, the planned route and the executed trajectory. Unknown space is treated as traversable; walls that appear a few metres ahead trigger braking and a replan without stopping the mission. Stack: PX4 SITL v1.17 (offboard control), ROS 2 Jazzy, Gazebo Harmonic, a persistent D* Lite 3D planner over revisioned raw occupancy, a CUDA MPPI trajectory optimizer, a stereo MSCKF written on Eigen feeding PX4's external-odometry interface, and a braking contract that bounds speed by what the cameras are guaranteed to have seen (6.4 m of confident depth). Release v0.4.0. Mean flight speed on this location: 1.5–1.8 m/s over five consecutive flights, no collisions, holds and replans included. Every flight also fails unless the drone's true position in the simulator is inside the 2.0 m capture radius when it declares arrival: an odometry drifts by metres, so without that check a drone can arrive perfectly in its own coordinates while standing somewhere else. Source code, docs and one-command setup (./scripts/bootstrap.sh): https://github.com/formiat/px4-ros2-d... Simulation only; not validated for real aircraft. *Contact*: Email: formiat95@gmail.com Telegram: @formiat