A CMU drone stack built for messy, dynamic airspace

What it does CERLAB’s framework is a full-stack autonomy suite for unmanned aerial vehicles. It bundles a lightweight Gazebo simulator, 3D mapping, global and local trajectory planners, a tracking controller, and an onboard detector for dynamic obstacles into a set of ROS packages. The stack targets three missions: navigating to a goal while avoiding collisions, exploring unknown environments, and inspecting targets with zig-zag trajectories. Everything is built to run against either a simulated drone or a real PX4-based airframe.

The interesting bit Most academic UAV repositories ship a planner and call it a day; this one ships the entire loop from sensor input to motor command, including a time-optimizer and dynamic-object tracker backed by recent ICRA and RA-L papers. The authors explicitly distinguish between a lightweight simulation branch and a PX4 branch that replicates real-flight behavior, which suggests they actually debugged the sim-to-real gap instead of ignoring it.

Key highlights

• Modular architecture with nine distinct packages, from uav_simulator and map_manager to trajectory_planner and tracking_controller.
• Handles dynamic obstacles, not just static maps, via the onboard_detector module.
• Supports both Gazebo/ROS simulation and PX4 hardware-in-the-loop and real flight.
• Published algorithms behind the stack: heuristic-based exploration, optimal time allocation, and RGB-D dynamic-object tracking.
• Targets Ubuntu 18.04/20.04 with ROS Melodic/Noetic.

Caveats

• Requires manually switching the autonomous_flight submodule between simulation and px4 branches and rebuilding, which is a friction point the README openly acknowledges.
• Real flight demands an external SLAM/VIO system and depth camera; those are inputs, not included modules.
• Locked to ROS1 (Melodic/Noetic); there is no mention of ROS2 or newer Ubuntu releases.

Verdict Robotics researchers and UAV builders who need a proven, end-to-end PX4 autonomy pipeline with dynamic obstacle support should take a close look. If you are looking for a ROS2-native stack or a plug-and-play commercial autopilot, this is not your framework.