Drone FC

Hover

A fully custom mini racing/stunt drone with ML-assisted balancing, automatic stunts, and some level of autonomous operation without costly cameras or GPS.

Why I built this

I wanted a challenge that forced me to learn PCB design and embedded firmware from scratch, but also let me apply the area I already know well - machine learning. Designing a drone from a blank schematic, writing every driver by hand, and then training a custom tiny ML model felt like the right way to do all of that at once. My goal was simply to push my first ever hardware project as far as I could, not to have a completely polished project.

What the project is

Hover is a tiny brushed micro quadcopter built entirely from custom hardware and firmware. There is no off-the-shelf flight controller, no commercial RC receiver, and no pre-written driver code. Everything was written/designed from scratch by me.

The system has two PCBs:

Flight Controller (FC): A 4-layer X-shaped PCB with a motor in each corner. It runs an STM32F103C8T6 at 72MHz, an ICM-42688-P IMU over SPI, a BK2423 2.4GHz RF transceiver, and four brushed motor drive circuits.

Remote Control (RC): A 2-layer PCB inside a 3D-printed enclosure. It uses the same STM32F103C8T6 and BK2423 RF ICs. It has two analog joysticks (ADC + DMA) for throttle, yaw, pitch, and roll, and eight buttons which are used for arming, auto take-off, auto land, return-to-home, a 360-degree flip trigger (one of my goals for this was to have some kind of smart one-button trick, and for this goal I settled on the 360 flip), and three spares for future use.

apart from that, this repository also includes:

ML Pipeline: A training and conversion pipeline for an onboard stabilisation model. Training data will be collected from real flight telemetry once the hardware is bought. The model will be converted to TFLite and dropped directly into the FC firmware, which already has the inference code in place and waiting.

CAD: Both the drone body and RC enclosure were designed in Onshape. The drone frame is a minimal 2-shell design with motor mounts, battery tray, and canopy, with both shells sandwiching the PCB in the middle. The RC enclosure uses the same 2-shell design with joystick cutouts, button extenders, and a battery bay below the PCB.

Firmware: Firmware created for both the FC and RC in STM32 CubeMX and CubeIDE, all firmware should work but is untested as I do not have the hardware right now. additionally, the ML inference file is writtenbut will not work until I create the .tflite model file.

BOM: A BOM generated from kicad for the FC & the RC and also other parts (excluding 3D printed parts and assorted screws). I will likely be purchasing all of these myself, and this is not completely final.

Build Photos:

Bill of Materials

FC

• Note that MPU-6050 is significantly cheaper locally, and that is why I chose to use it despite cheaper alternatives

RC

Other

• This list covers physical assembly components not part of either PCB
• It is not completely final - I made this off the top of my head, there are probably other components I forgot to identify
• It does not include miscellaneous screws or 3D printed parts