Is There Any AI That Can Tune A Betaflight Drone? Could You Train One? - FPV Questions

Um, Clipsy asks, "Is there any AI that

can tune from Blackbox info?" I actually

addressed that

recently on the Joshua Bardwell live

stream clips

channel. Uh, that's where Blunty clips

out uh, segments from the live stream

and uploads them for you to enjoy.

And I've

recently, let's just search. Can chat

GBT tune your drone?

No, it

cannot. Well, if you want to go watch

that clip to get the details, you can.

It's on the Joshua Bardwell Live Stream

Clips channel. The short version is I I

think the answer is no

because I am pretty sure that in order

to do PID

tuning

you you have to be able to parse the

blackbox data in ways that I don't think

Chat GPT can

specifically in order to take the gyro

data and convert it into a frequency

plot. You have to do a fast Forier

transform which is a mathematical

function that does that thing. And chat

GPT can't do a fast Forier

transform. So when chat GPT tells you to

do something with your filters and then

it works for you, I think it's just

hallucinating and by chance it got the

answer right.

And I also don't think that chat GPT has

the ability to calculate like step

response like PID toolbox

does.

So if if you're handing chat GPT a

blackbox log, I don't think it knows how

to parse a blackbox

log. It's possible that if you give it a

CSV data, it will be able to parse the

CSV data. And you can convert a blackbox

log to CSV, but I still don't think it

actually has the core logic that it

needs to be able to interpret and make

recommendations on the data. And people,

it drives me crazy. And if I'm wrong

about this, I will happily admit that

I'm wrong. Chat TPT can do some cool

[ __ ] you know, but I just don't think

it

can parse blackbox data. Just to be

clear, you're also people in chat maybe

not, but you're using chat GPT as a

blanket term for AI LLM in general.

Yeah, LLMs in general. Is there any LLM

that can do a fast for transform? Can

any LLM? Well, I mean, any of them could

do it with code, right? Like if you're

in cursor and you prompted one, like you

could get a fast for

transform, you know what I mean? Like

through code. If you have an LLM, if you

had an LLM that could access a Python

function that could do a fast fora

transform. Yes. Yeah. I mean, it would

have to know that it needed to do that.

Like you could go and cursor prompt the

AI to get you something that could tune

a drone with the thing and explain. You

know what I mean? That's the sort of the

idea. But then you would have to work,

right? That's not what they do. So here,

for example, please write me a Python

function to calculate FFT.

Uh a fast for a um an AI could do that.

But what people are doing is

they're what they're doing is they're

just

saying I dumped a blackbox log. I can't

find I can't find an example. They're

just saying, "I dumped a blackbox log

into chat GPT." And then chat GPT goes,

"Cool. I will help you tune your

blackbox log. We will get maximum step

response and you know, properly tune

your filters." And it just says a bunch

of

[ __ ] And then it's like based on

what I see in your blackbox log, I

recommend that you increase your P gain.

And it's like, you know what? Like I'll

put on a white lab coat and a

stethoscope and hold a clipboard and

I'll say, "Cool. I've looked at your

I've looked at your blood test results

and it seems that your cholesterol is

132. I recommend that you get that

number up. A good cholesterol is between

187 and 221. Don't I sound

confident? I I'm completely just talking

talking out my ass. And I think that

when when chat GPT or any AI pretends to

be like it knows that terms like P gain,

D gain, filters, step response,

overshoot, oscillation. It knows that

these are terms that are associated with

blackbox logging and it makes sentences

that sound convincing and makes

recommendations which you then follow

and maybe they work, but it's not

because it actually understood what was

in your blackbox log. I don't think it

can possibly know

that. So

yeah, and this also isn't to mean that

somebody can't eventually train

something to do this specifically. Like

one of the things we're seeing now is

agentic. Somebody mentioned that in the

chat, agentic AI, where you have like

one I think that's how DeepC handled it.

You have one big LLM model, but it's

it's it's basically asking individual

agents that are good at certain jobs. So

it'll have a agent that's good at math.

And so if it has math in its thing,

it'll go ask the math agent to solve it

for it and bring it back. So the idea is

that you would have specific agents that

understand these pieces and it can't

really get lost as easy because it knows

to ask the thing who doesn't have as

much context like who doesn't have

contact. But but here's the problem with

that.

It ha there has the training set, the

training data has to include

examples that let the large language

model learn what right looks like. Does

that make sense? Of course. Well, the

that's how LLMs work. And obviously a a

treatise on how LLMs work is not I'm not

qualified to give it. But the short

version is that you feed the LLM a lot

of

data

that is

correct or that from which you want to

draw

inferences and then you ask the LLM

questions about the data and it tells

you and talks to you about the data but

garbage in garbage

out. Did any LLM get trained on blackbox

logs?

Can it even parse a blackbox log? If I

hand it a blackbox log from Betaflight,

it's just a bunch of ones and zeros.

Does it know how to interpret that? How

would it even know?

And if it Yeah, you build an interpreter

like that. You would have to build it. I

mean, that's you would have to do but no

one's done that. My point is no one's

done that.

Yeah. So you would have to number one

you would have to build an

interpreter so that it could in

basically you would have to build pit

toolbox into chat GPT or give it an API

that let it access PID toolbox so that

it could look at the blackbox log and go

okay your frequency response is X your

sorry you know your your step response

is Y here are your current PIDs and

based on that and then it's got to make

a recommendation. So again, you would

have to train it on when P is P goes up,

here's what happens. When P goes down,

here's what happens. And here is what

we're looking for. We're looking for

this. This is our end state that we

want. And no one's done that. No one's

done that.

I think the alternative would be, is

there enough data out there? If it comes

all the forums and all the things and if

it get transcripts of YouTube videos

like things like that, right? like is

there enough data out there to tell you

how to tune a drone or like you you were

saying you need to specifically tune

something and then like you said there's

also in a blackbox interpreter so you

need the data out of it. So uh you know

that's something they're doing more more

though is like now their image