ChatGPT for settings

General
1

For those who haven’t tried ChatGPT for Ardupilot settings, you should. I was pleasantly surprised at the responses when I asked questions like “what parameter changes should I make in order to maximize range on my quad with 13” props and ….”. Although I don’t accept anything without testing, I find the results to be surprising accurate.

2

could you give examples??

3

It is okay to use it as long as you treat it like a glorified web search and don’t blindly accept everything it says.

You must:

  • Understand parameters before changing them.
  • Use it to get ideas, not final settings.
  • Always consider that some ideas will be hallucinations.
  • Learn the underlying concepts and modes of operation.

Always validate any recommended changes using real data, logs, and controlled testing.

4 Likes
4

I agree. I was trying to get a little more range from my plane. AI mentioned two parameters I hadn’t thought of, as well as a trick that I would never have found on my own.

5

Would be nice if you could explain that trick.

1 Like
6

you have posted twice with zero information on anything other than “I asked chat gpt and liked what it said”

2 Likes
7

Here is some actual output from ChatGPT:

I was getting too much overshoot on turns, so I asked what I should try first.

If you’re flying fast in AUTO:

Check:

TRIM_ARSPD_CM
AIRSPEED_CRUISE

Reducing cruise speed slightly reduces overshoot dramatically.

:white_check_mark: :five: Optional: Waypoint Turn Behavior

If you’re using straight-line waypoint logic:

:wrench: WP_RADIUS

If too small, plane aggressively tries to hit exact point → overshoot.

Typical values:

WP_RADIUS = 60–100 meters (for fast planes)

:microscope: What I Would Adjust First

In order:

  1. Reduce NAVL1_PERIOD from 20 → 17

  2. Increase LIM_ROLL_CD if below 50°

  3. Slightly increase RLL2SRV_P

  4. Verify cruise airspeed isn’t excessive

:warning: Important Question

Does it:

  • Overshoot then smoothly correct?

  • Or oscillate back and forth?

Those are two different tuning problems.

If you tell me:

  • Airspeed

  • Wing span

  • Typical bank limit

  • Current NAVL1_PERIOD

8

As it is happening pretty often with what this generation calls “AI”, the outputs of large language models seem plausible, logical and consistent… unless you actually understand the matter, and if you do, you see how much pointless, wrong and plainly dangerous LLM’s advice often is.

One of my “clients”, who for various reasons finds it difficult to ask me for help every time help is needed, has once run a diagnostic session for a particular machine using ChatGPT and logs produced in few flights. While it “can read” these logs - in that it indicated some features that were definitely present in these logs - it made complete and severe misjudgement of root causes of the undesired behaviour, which, if implemented, would result in an outright dangerous tuning of the copter.

For instance, for a lightweight 10" quad with quite powerful power train, including a 6S battery, it suggested getting angular accelerations from their working values of around 300K to 80K, while increasing PIDs from working values of around 0.05/0.002 straight to 0.12/0.007. If implemented, this could have burned the motors by a D-term excitation, possibly combined with angular control so sluggish that damage to the surroundings was very likely. This was all very strongly “motivated”, at times using very persuasive language, referring to almost non-existent dependency of rate outputs from the pilot inputs - which was solely due to the fact that the inputs were very tame compared to what that copter could handle.

Part of the reason is, as usual, that ChatGPT and other LLMs like this do not differentiate truth from falsity, facts from opinions, and while it has “read” this entire forum several times (as in: every text from this forum has likely appeared in its training dataset in multiple copies), it has exactly zero capabilities to get any understanding, let alone choosing correctly from multiple different answers - and this forum definitely has seen much more wrong answers than the correct ones.

And it bears no responsibility. If your machine crashes, and you tell it so, it will of course write tons of apologies and will find numerous random reasons for the problem, most again plainly wrong.

1 Like
9

Agree. The answers presented here are pretty plausible, but you’d be better off using this forum or Discord to get expert advice rather than applying LLM suggestions, especially for the somewhat rudimentary stuff shown so far.

1 Like
10

I posted an extremely small part. AI was very helpful in finding parameters that I might have otherwise missed. My goal is to make a plane fly a total of 82KM. That isn’t trivial.

11

That is exactly why you should likely seek expert or even paid tuning help.

I understand you’ve been satisfied with what you’ve gotten from the LLM - that’s not necessarily wrong, but it’s highly likely that it’s incomplete or in error in some or even many respects.

3 Likes