3 ms·
Show HN: I Built an AI Drone Controller
Hey HN! I'm someone who both flies and programs drones. I've often struggled to make my drones perform perfectly, especially when the environment changes like a gust of wind.
To solve this, I’ve trained ML Models to replace the Control Loops currently in charge of controlling most drones. These models perform better than Control Loops, require no tuning, and retain this performance in many more situations than Control Loops.
Requiring no tuning also means that the same model can be used for multiple different drones/configurations, so it works out of the box.
A high level model is also trained to perform obstacle avoidance and navigational pathfinding.
Check out the demo: https://youtu.be/fWgD7tjZugY https://youtu.be/fWgD7tjZugY
I'd love to hear any feedback or questions you might have! Thanks :)
- m1keAttack 3y agoThis looks amazing! I’ve spent so much money crashing my drone while trying to tune it. Does this sit on top of an existing flight controller firmware? Or does it replace whatever firmware we are using?
- hagen6835 3y agoRight now, we build it into the PX4 Firmware, so it's essentially a PX4 plugin. If you use something like ardupilot or dRehmFlight, we would have to build another wrapper for that (which we are working on now). We do eventually want to build our own firmware eventually that would entirely replace the functionality of whatever firmware we are replacing.
- filoleg 3y agoI have no domain experience myself, so just wanted to make sure I understood it right. Is it similar to having an ECU in multiple different cars with entirely different engines and transmission and everything else, running the exact same AI model that wasn’t custom-tuned, and still delivering better performance than the default ECU? Does this analogy sound right?
- hagen6835 3y agoYeah, exactly. Since it was trained to fly in a variety of situations, it is able to recognize the situation it is currently in and adapt accordingly.