Autonomous Drone Platform
A vision-guided quadrotor for GPS-denied flight, built around a Jetson Orin Nano and an Orbbec Gemini 336 depth camera on a laser-cut carbon fibre airframe. Final year project, in progress.
- Jetson Orin Nano
- ROS 2
- Orbbec Gemini 336
- C++
- Python
- CAD
The problem
Autonomous flight in cluttered, GPS-denied spaces only works when the whole stack does. It needs a depth sensor that returns usable geometry indoors, and enough onboard compute to close the perception loop without a ground station. It also needs an airframe stiff enough that vibration doesn't destroy the state estimate.
This is my final year project - building that platform end to end rather than buying it.
Approach
I took the Stanford MSL TrajBridge drone as the hardware reference design and worked outward from the compute and sensing I already had.
Perception and compute. An Orbbec Gemini 336 depth camera paired with an NVIDIA Jetson Orin Nano for onboard inference. The Orin Nano is the pivot of the design. It can run depth processing and a navigation policy onboard, so the aircraft doesn't depend on a radio link to stay alive.
Airframe. I originally planned a 3D-printed frame and abandoned it after working the numbers: for the stiffness required it came out both too heavy and too weak compared to carbon fibre. The design moved to laser-cut carbon fibre sheet for the primary structure, keeping 3D printing for non-structural parts - the Nylon 12 prop ducts, mounts and sensor brackets - where its geometric freedom actually pays for itself.
Propulsion. Four 2205 motors on 5 in three-blade props inside the ducts, on a 222 mm diagonal, sized against the mass budget the carbon frame and the Orin Nano impose.
Explore the airframe
I built an interactive schematic of this aircraft - orbit it, and select any component to see its specs, what it's wired to, and whether it actually exists yet. The geometry is the aircraft's own CAD, and each marker sits on the part it names.
Where it stands
In progress. The full airframe is modelled in CAD, and the compute and sensing selection is settled; the current work is sourcing propulsion and integration.
What I took from it
The frame decision was the useful lesson. The 3D-printed version was the obvious path, and the analysis said no. Convenient manufacturing is not the same as appropriate manufacturing, and on an airframe the mass budget arbitrates every argument.