Dynamic Soaring Controller

A defensive publication of a flight-control concept I had originally drafted as a provisional patent: a learned dynamic-soaring controller for large electric and hybrid-electric aircraft, anchored to a hardware safety supervisor and shaped by a passenger-comfort reward.

I am not a flight controls engineer. I am a 23-year-old machinist from Wichita who has spent a lot of time staring at albatross videos and at electric aircraft roadmaps. Everything in this repo is an idea, not a result. I am putting it in the open so the people who actually build these things can pick at it.

Plain English — what is this, really?

A wandering albatross can fly tens of thousands of miles a year on almost no calories. It does it by climbing into the wind gradient above the ocean and diving back down through it, harvesting a little energy on each loop. This is called dynamic soaring.

People have already taught small UAVs to do this. The piece I have not seen done is doing it on a big aircraft — one with passengers, one where you cannot just yank 4 g of roll because someone will spill their coffee, and one where if you ever clip a wave it is a disaster, not an oops.

The idea here is a control stack that does three honest things at once:

1. Learned policy — a neural-network or other ML controller that learns dynamic-soaring maneuvers in a physics simulator.
2. Comfort-shaped reward — the simulator penalizes the policy any time it produces accelerations in the 0.1–1.0 Hz band that humans actually feel, so the policy learns to be smooth, not aerobatic.
3. Hardware safety supervisor — a small dedicated circuit, separate from the main processor running the ML model, that watches a forward-looking FMCW wave-clearance radar and yanks pitch up if the aircraft ever drifts under a minimum altitude. The ML model never gets the last word on safety.

That third piece is the one I think is the real story. Putting safety in a hardware override instead of trusting the learned model is what makes this a thing you might actually fly people in.

A secondary version of the same architecture applies upstairs — using forecast data to ride jet-stream shear at cruise altitude.

What this is not

• This is not a finished flight controller. I have not flown it. I have not even built the simulator.
• This is not a patent claim. I am publishing this material publicly so it becomes prior art and stays available to anyone who wants to use it.
• This is not novel in the sense that "nobody has flown dynamic soaring." UAV dynamic soaring is a real research area, going back to Sachs (2002) and Bower (Stanford). NASA holds a thermal-soaring patent (Allen, US 7,431,243). What I have not been able to find in the literature is the specific combination of (a) large passenger-class aircraft with high wing loading, (b) a learned policy trained against a passenger-comfort reward in the 0.1–1.0 Hz band, (c) a hardware-isolated safety supervisor driven by an FMCW wave-clearance radar, and (d) an energy-state aggressiveness scalar that ties soaring effort to SOC, solar output, and fuel cell output. If you find that combination already published, please open an issue and link it.

What is in this repo

dynamic-soaring-controller/
├── README.md                       <- this file
├── LICENSE                         <- Apache 2.0
└── docs/
    ├── DEFENSIVE_PUBLICATION.md    <- the full provisional-style writeup
    └── PLAIN_ENGLISH.md            <- a longer "explain it like I'm 12" version

What I am asking for

I am putting this in the open because I would rather see somebody build a real version of it than own a paper version that never goes anywhere. If you are a flight-controls engineer, a controls grad student, an RL researcher, or anyone working on electric aircraft, here is what would help:

• Tell me what is wrong with the architecture.
• Point me at prior art I missed, especially anything that already combines hardware-isolated safety overrides with learned dynamic-soaring policies on large aircraft.
• If you actually fly RL controllers, tell me what assumptions in the comfort-shaped reward are naive.
• If you work on electric aircraft (Eviation, Regent, Heart, Joby, Beta, anyone), tell me what you think the realistic energy-savings ceiling is for a large aircraft over open water.

Open an issue. Be blunt. I would rather hear "this is wrong because X" than nothing.

License

Apache License 2.0. Use it, fork it, build on it, ship it. If something here ends up in a real product, all I ask is a mention somewhere in the credits and that the chain of open-source attribution stays intact.

— Khai Bustos, Wichita, Kansas