From Meteor75 Pro to Meteor75 Pro II: a week of fighting resonances

Craft name Snake. It started life as a Meteor75 Pro, and it is now a Meteor75 Pro II — frame and canopy ordered off AliExpress, everything expensive carried straight over. Same Matrix 1S 3-in-1 FC. Same narrow-FOV DJI O4 air unit. New shell, old guts, and by the time I was done, 169 flights and 15,574 seconds of logs to argue with.

The plan was a fifteen-minute swap. What I got was a week of chasing a resonance, three retractions — seven of them, itemised later — one clean hypothesis that was completely wrong, one tuning change I had to revert, and a metric that lied to me for several rounds before I noticed.

The thesis of the whole post: the canopy that fixed my jello is the canopy the flight controller now has to fight. Decoupling the camera from the frame is good. Decoupling it softly is not free.

The build, and the mismatch that matters

Left: the old Pro frame and canopy, stripped. Right: the Pro II, built up. Same flight controller, same air unit, same motors — everything that changed is structural.

• Frame + canopy: Meteor75 Pro II, AliExpress parts
• Guts: carried over from the Meteor75 Pro — same Matrix 1S 3-in-1 FC, same narrow-FOV DJI O4 air unit
• FC target BETAFPVG473 (STM32G473), manufacturer id BEFH
• Betaflight 4.5.1 (Dec 11 2025, 77d01ba3b)
• 1S LiHV — vbat_max_cell_voltage = 435, auto_profile_cell_count = 1
• DSHOT300, dshot_bidir = ON, motor_poles = 12
• 3.2 kHz gyro and PID loop — looptime 312, pid_process_denom 1
• blackbox_sample_rate = 1/2 → 1582 Hz logging, 791 Hz Nyquist
• Digital VTX over MSP DisplayPort on serial 3
• yaw_motors_reversed = ON (props out)

The central part, which I did not think about at all while clicking "buy": the Pro II canopy was redesigned around the O4 Wide. Snake runs the narrow-FOV O4, so the canopy is not carrying the mass it was drawn around and the FC/canopy interface is not the pairing the frame was designed for. I was building a hybrid and calling it an upgrade.

Two things I checked rather than assumed. motor_poles = 12 is a setting, not a measurement, so I validated it: measured dominant roll frequency over computed 1× came to 1.008–1.020. At a physical pole count of 14 that ratio would land near 1.17. The RPM filter was targeting the right frequency all along.

And my PID sliders were doing nothing. simplified_pids_mode = OFF in profile 0 means the configured slider values (master multiplier 120, d_gain 120, pi_gain 120) were inactive. Profile 0 flew stock Betaflight 4.5 defaults the whole time: roll 45/80/40, pitch 47/84/46, yaw 45/80/0. Worth knowing before you spend an evening theorising about your tune.

The symptom

"Flying in the yard with some wind, I got huge vibrations."

First log, old props. Roll axis pre-filter HF energy (80–780 Hz) came in at 68.5 °/s RMS. Pitch: 8.0. Yaw: 11.4. That is an 8.6 : 1 roll-to-pitch ratio, which is not a noise problem, that is a single-axis mechanical problem wearing a noise costume.

Post-filter the same axis read 1.38 °/s — the RPM filter carrying roughly 34 dB and politely hiding a large mechanical fault from the flight controller. The quad flew fine. The gyro was screaming.

The harmonic structure named the fault type: the 1× to 2× ratio was around 200:1 (53:1 to 212:1 depending on the motor), which is textbook mass imbalance. A bent blade or real aerodynamic loading puts energy into the higher harmonics; this put essentially none.

Caveat I am not going to quietly drop: at roughly 341 Hz the 3rd harmonic lands at 1023 Hz, above this log's 791 Hz Nyquist, so blade-pass content could not be assessed at all. The 2× at ~682 Hz was in range, clean, and the diagnostic one — so the conclusion holds on 2×, not on a full harmonic picture.

More wind made it better, which should be impossible

My first instinct was that this was a wind problem. It said so right there in my own note. So I compared sections at matched prop frequency (330–350 Hz), to hold the resonance constant and let only the air change.

corr(turbulence, vibration) at fixed RPM = −0.584.

More wind, less vibration. Dead-still indoor air was the worst case I could produce.

That is the single most useful result of the week, because it kills the obvious explanation on day one and because the reason it happens turns out to be the mechanism. It takes a few more sections to earn.

Two things my config was getting wrong

Before chasing physics I read my own filter settings properly, which I should have done first:

1dyn_notch_count   = 1     (default 3)
2dyn_notch_q       = 400   (very narrow)
3dyn_notch_min_hz  = 150
4dyn_notch_max_hz  = 350   <-- BELOW the measured 342-357 Hz peak
5gyro_lpf1_static_hz   = 0 (LPF1 fully disabled)
6gyro_lpf1_dyn_min_hz  = 0

One notch, made needle-thin by q = 400, with a ceiling below the actual peak — the one filter aimed at this problem physically could not reach it. LPF1 was off entirely. The fix:

1set dyn_notch_count = 3
2set dyn_notch_q = 300
3set dyn_notch_min_hz = 100
4set dyn_notch_max_hz = 600
5set gyro_lpf1_dyn_min_hz = 250

Measured at matched prop RPM:

Pre-filter was unchanged, which is the whole point: filters protect the loop, they do not fix the airframe. The quad shook just as hard afterwards. The flight controller simply stopped reacting to it.

The measurement bar — the number I should have established first

Everything after this depends on one boring question: how big does a change have to be before I am allowed to call it real? So I measured the scatter of pre-filter roll HF RMS within a single flight, at fixed RPM, and treated that as my noise floor:

1CV = 9.0%,  max/min = 1.38   (n = 21 windows of 3 s)
2corr with pack voltage      = +0.04
3corr with time/temperature  = -0.05

Any change smaller than about ±10% is indistinguishable from noise. Not "probably noise" — indistinguishable. Not pack sag, not thermal drift; both correlations are flat. It is just how much this measurement wanders when nothing changes, and it killed several conclusions I wanted to keep later in the week. Establish the noise floor before you believe any result — especially one you like.

Props: the first real mechanical win

New props changed three things at once — bad experimental hygiene, good evening:

• RPM-per-output spread across the four motors collapsed from 9.2 to 4.4 percentage points
• 1× amplitudes evened out — m1 108.7 → 56.7 °/s, m4 107.1 → 56.8
• hover prop frequency dropped 330 → 308 Hz

Outdoor, full RPM sweep, same airframe, so what changes here is the forcing:

Cut at 425 Hz on purpose: the 450 and 475 Hz bins carry 1.1–3.0 s of dwell against 32–53 s in the bins that matter, and a 4 s excursion through a resonance cannot build the amplitude that 50 s parked on it does. Every bin shown clears 4 s on both flights.

−45% at the peak, −56% at 350–375 Hz. Fixed-band energy across 325–365 Hz went 1185 → 263 — a 78% cut.

Note where the curves start: at 275 Hz they are identical at 42 °/s. Below the resonance the props make no measurable difference at all. Everything they bought, they bought inside the band — the first hint that this was never really a prop-balance story.

At this point I thought props and a notch config had solved it. I had not even correctly described what the problem was.

Why it only shook sometimes — and a clean hypothesis that was wrong

The observation that cracked it is one I nearly ignored, because I had written it down and then filed it as vague: the shaking is not always present, only in some orientations relative to wind.

Intermittent. Orientation-dependent. My first idea was beat frequencies: four motors at 343 / 313 / 337 / 332 Hz predict beats at 5.2, 6.1, 11.3, 19.7, 24.9 and 31.0 Hz, right in the band where I could see the airframe moving. Clean, testable, satisfying, and wrong:

1coherence(beat envelope, visible 8-45 Hz motion) = 0.019 mean, 0.063 max
2corr(RPM spread, envelope)                       = -0.287    (wrong direction)
3measured modulation 1.9 Hz vs nearest predicted pair 5.2 Hz

Coherence of 0.019 is not a weak signal, it is no signal, and the RPM-spread correlation came out negative — the opposite of what a beat model requires. Dead in one afternoon.

What did predict the shake was duller:

And the dose-response, about as textbook as field data gets:

It doubles. Count how many props sit inside the resonance window and you can predict the shake.

That explains the intermittency, the orientation dependence and the backwards wind correlation in one go. Wind redistributes thrust between the corners, shifting individual motor RPMs by 20–40 Hz and sliding them in and out of the window — gusts scatter the props off the resonance. Indoors the quad hovers rock-steady and parks all four of them on it for as long as you let it. Still air is the worst case because still air is the most precise. That sentence comes back every time I compare an indoor flight to an outdoor one.

It also reframes the prop win:

The old props hovered dead inside the resonance band, three hertz of margin. Less imbalance was the smaller part of the win; moving the operating point off the resonance was the larger part. I had done the right thing for a reason I did not understand.

flowchart TD
    A["Symptom: huge vibrations in wind<br/>roll 68.5, pitch 8.0 — 8.6:1"] --> B{"Is it the wind?"}
    B -->|"corr = -0.584<br/>still air is WORST"| C["Wind ruled out<br/>(and inverted)"]
    C --> D{"Could the loop even see it?"}
    D -->|"dyn_notch_max_hz 350<br/>peak at 342-357"| E["Filter gap:<br/>notch could not reach the peak"]
    E --> F["Fix filters: -70.6% post-filter<br/>pre-filter UNCHANGED"]
    F --> G{"So what is shaking?"}
    G -->|"1x:2x = 200:1"| H["Mass imbalance<br/>-> new props"]
    H --> I["-78% band energy<br/>hover 330 -> 308 Hz"]
    I --> J{"Why intermittent?"}
    J -->|"coherence 0.019<br/>RPM-spread corr wrong sign"| K["Beat hypothesis DEAD"]
    J -->|"r = +0.652"| L["Resonance proximity<br/>dose-response doubles"]
    L --> M["Wind shifts RPM 20-40 Hz<br/>in and out of 325-365 Hz"]
    M --> N{"Loop cannot reach 320-345 Hz.<br/>What is left?"}
    N --> O["The MOUNT<br/>peak gone, dose-response flat"]
    O --> P["Mechanism unresolved:<br/>stiffness / modal mass / damping<br/>move together"]

    style K fill:#915d52,color:#fff
    style L fill:#244d68,color:#fff
    style O fill:#244d68,color:#fff
    style P fill:#bd9361,color:#000

Two problems, not one — and the catch in "Gyroflow fixes it"

This distinction took most of the week to nail down, and it decides what software can save you from.

(a) The ~320–345 Hz structural mode. Roll-dominant, 8:1. This is the jello source. It sits an order of magnitude above the control loop's usable bandwidth of 20–40 Hz. No PID change, no TPA setting, no filter tweak can touch it. Filters stop it reaching the loop; they cannot stop the airframe shaking. And neither Gyroflow nor RockSteady can remove jello — it is intra-frame distortion, the damage is inside the rolling shutter before any stabiliser sees the image.

(b) Broadband 10–25 Hz turbulence following. Measured Q ≈ 1.9–2.2. Peak 15.8–17.8 Hz on roll, 10.6–12.9 Hz on pitch, amplitude 4.4–5.3 °/s. A control-loop limit cycle would show Q = 10–100; Q ≈ 2 is a lightly-damped airframe genuinely being pushed around by turbulent air.

There is a genuinely sharp mode in there — 48.5 Hz at Q = 83.7 — with an amplitude of 0.24 °/s, i.e. completely irrelevant. High Q is not the same as important.

Where does the motion you can actually see live?

Filters alone took the high band 1.68 → 0.38, props took it further: −66% at 1–5 Hz, −85% up high. Note the ratio — roughly five times more energy sits in the Gyroflow-correctable band than up where a rolling shutter turns vibration into jello. Which is why the footage looked acceptable while the gyro was screaming.

The thing I am actually fighting. Top trace is raw gyro: a continuous band that swells and fades rather than a clean line. Everything below it - filtered gyro, P, D, PID sums - is flat, which is the filters doing their job. None of that helps the camera.

On its own that trace is just a number in a log. The problem is what happens next: under some conditions the loop reacts to it, drives the motors with it, and the airframe genuinely moves. Then it stops being a gyro reading and starts being shake in the video feed — and not only in heavy wind, in conditions I would call mild.

Now the catch, and the most useful practical thing I learned all week:

Video stabilisation only rescues this if there is plenty of light.

On a bright day exposure times are short, each frame is crisp, the shake shows up as frame-to-frame displacement, and Gyroflow can re-align frames and take it out. On a cloudy day the camera holds the shutter open longer. Now the shake happens during each exposure rather than between frames, and it is recorded as motion blur baked into the pixels. Stabilisation can align a blurred frame perfectly and it is still blurred. The whole clip is soft.

So the comfortable framing — jello unfixable, low-frequency shake fixable — is too generous. The honest version has three tiers:

Two of the three are unrecoverable, and which one you get is decided by the weather rather than by anything in the tune. That is why I kept after the mechanical side long after the flight controller had stopped complaining.

A tuning experiment that failed and got reverted

The D-term lagged the error by 16.4 ms in the 8–45 Hz band, most of a half-cycle at 17 Hz, so raising dterm_lpf1_static_hz from 75 to 90 looked like free money. Matched indoor hover, same props, 307 vs 309 Hz:

It bought 1.9 ms of lag reduction for 370% more motor jitter, and the spectrum was worse at every frequency from 2 to 400 Hz. Reverted.

Airmode went on in the same session (feature mask delta exactly 4194304 in the log) and stayed — 3.3 s below 1250 throttle at minimum motor output 201, no authority dropout. Confound: two variables changed at once, so the 14 Hz growth cannot be cleanly split between the filter and airmode. The other four rows are large enough to survive that; the 14 Hz number is not clean.

Why I could not measure my own step response

I tried repeatedly to get a real step response out of these logs and was blocked by the input:

1setpoint energy: roll 95% below 1.7 Hz | pitch 1.4 Hz | yaw 1.5 Hz
2hard stick reversals: 0
3slew events >4000 deg/s^2: 3

A quad's loop lives at 20–40 Hz. Smooth continuous rolls contain no high-frequency content, so the step response is input-bandwidth limited, not quad-limited. The "173 ms rise time" I computed early on was a faithful measurement of my sticks.

One flight with 39 hard reversals and 26 sharp slews finally gave a real number: roll overshoot +10.4% at 133 ms, rise(90%) 77.7 ms, 50% delay 32.9 ms — n = 6 steps, because the log ended in a 9.6 G crash. Indicative, not settled.

The bad motor that turned out to be air

For most of the week one motor looked guilty:

1m2 RPM-per-output:  -4.2% to -6.1%    (worst in EVERY log)
2m1 hover output:    +6.7% to +11.1%   (works hardest, and the ONLY motor clipping)

m1 clipped 0.789% of frames while m2 and m3 sat at exactly 0.000%, and the shake was 1.59× worse with motors near the ceiling. Diagnosis: a draggy bearing on m2 and an overworked m1. Two hardware faults, both confidently called.

Then I rotated the canopy 180° and the ordering reversed:

1before rotation:  m2 = -4.2% to -6.1%   (worst)
2after rotation:   m2 = +4.3% to +8.0%   (freest)

A motor defect cannot flip sign when you rotate a canopy. The pattern is aerodynamic — the canopy shadows whichever props sit under it. Both diagnoses retracted. Installation, not hardware, and I only found out because I changed something unrelated and kept measuring.

The rotation did real work on CoG:

Hardest-working motor moved from m1 to m3/m4, and m1's clipping went 0.812% → 0.000%. The rotation alone took the front/rear pair split from +9.5% to +3.6%.

Two notes on scope, because these numbers are easy to mis-stack. The +12.5% on m1 is the 15:53 outdoor flight specifically, while the +6.7% to +11.1% range above spans the 14:26, 15:20 and 16:28 logs — one flight versus a range across three, neither superseding the other. And the rotation and the foam are separate interventions whose CoG results do not chain: rotation moved the pair split +9.5% → +3.6%, the foam later and independently moved it +3.4% → +2.0%. Reading those as one run from +9.5% to +2.0% would be wrong.

The mod is not mine — rotating the canopy 180° comes from Oscar Liang's Improvements You Can Make section.

A battery weighed from a log file

Small side-quest, included because I enjoyed it. Two packs flown back to back; hover RPM is a valid mass proxy at fixed prop and config:

1log1: airborne 70 s, hover 330 Hz, 966 indicated charge
2log2: airborne 95 s, hover 340 Hz, 1585 indicated charge
3hover RPM ratio 1.0612 -> mass ratio 1.126 -> log2 is 12.6% heavier

Identified purely from the log, with no other input about which pack was which.

The packs are also the practical reason the canopy went round the other way: rotated, it gives better mass distribution with the LAVA 2 680 mAh batteries I actually fly, so the front/rear split halving was design intent rather than luck. Those packs buy about 3 minutes ripping, 5–6 cruising. Worth holding next to the heavy-versus-light thread further down — the heavier pack bought 36% more airborne time and 4× more motor clipping.

The mount is the lever, not the tune

The canopy the whole argument is about — drawn around the O4 Wide, carrying a narrow-FOV O4. It isolates the camera far better than the old one. It also gave the flight controller something to fight.

The loop cannot reach 320–345 Hz. The props were already good. That leaves the structure, and the thesis of the post: frame-and-canopy decoupling is good and bad at the same time.

The old canopy coupled too much vibration into the camera: jello, unfixable downstream. The new one is far better isolated, so what the camera still sees is low-frequency and Gyroflow-shaped — conditionally. But that same decoupling created a soft, lightly-damped path between the FC/canopy assembly and the frame, and the FC now fights the canopy. In higher winds it loses, because wind shifts motor RPMs into the resonance window and the mode gets driven.

I spent the first half of the week adjusting a 20–40 Hz control loop in the hope of influencing a 320–345 Hz structural mode. That was never going to work, and it took a dose-response curve to convince me. What follows is five mount configurations in the order I flew them, with the conclusion moving twice.

One: a large foam pad, and the amplification collapsed

Stiff foam between FC and VTX, stretching the gummy-ball mounts and stiffening the canopy fixation. Same pack (hover 345 vs 347 Hz), zero config changes — a clean mechanical A/B.

The dose-response that had defined the entire problem collapsed:

Vibration used to climb 45–63% as motors entered the band. Now it is flat: motors sitting in the resonance band stopped mattering, which is a better outcome than making them smaller.

The resonance curve says the same thing:

The result is the disappearance of the peak, not a reduction in its height, and the distinction is deliberate. Before the foam there is an unmistakable amplification peak at 48.8 °/s. After, there is no peak: the curve sits between 25 and 30 °/s across the whole 250–425 Hz sweep and the "maximum" is wherever the noise lands that run. Quoting a single after-peak number invites a percentage that is really a resonance compared against a flat line, so I am not quoting one. The curve stopped having a shape. That is the finding.

And the energy did not vanish, it moved:

Caveat: throttle p99 was 1751 versus 1968, so part of that zero-clipping result is me flying less aggressively — the weakest row in the table, and it should be read as such. The pair-split row is the foam result, independent of the canopy-rotation result earlier.

Two: TPU in the gummies, because foam cooks the board

The foam worked, but it is a blanket over the hot part of the board, so it came out. What replaced it was two changes in the same session:

1. The VTX mounted directly to the canopy, silicone grommets removed. That deletes a compliant element from the path between the air unit's mass and the canopy — canopy and VTX are now effectively one body.
2. TPU filament inserted inside the gummy balls, raising their stiffness substantially and stiffening the FC-to-frame path.

Two stiffness increases, on two different load paths, at once. So whatever the numbers show, I cannot split the credit between them — a self-inflicted attribution problem, and the honest move is to label it rather than pick a winner.

TPU filament pushed inside the rubber gummies. The red circle marks one of them. Two jobs, not one: stiffer coupling, and a canopy far less likely to part company with the frame.

The TPU's second job needs no measurement: with filament inside them the gummies are far less inclined to separate, worth having on its own for a whoop that spends its life bouncing off door frames. Oscar Liang uses glue; I used filament, because glue is a one-way door and filament comes back out — which matters when the point of the exercise is A/B testing the mount.

The scoring plan was written down before the flight, because I no longer trust a comparison I designed after seeing the data. The primary criterion was the motors-in-band dose-response staying flat.

It stayed flat. 84 s of clean indoor hover, second arm, zero impacts, 0 config changes. On the resonance curve only one bin is trustworthy — 79.6 s of dwell at 300–325 Hz against 0.5–1.8 s everywhere else — so I am plotting that point alone rather than drawing a line through noise:

39 °/s, between the 49 of no foam and the 26 of foam — except the two curves are outdoor and that point is indoor, which per the first finding in this post is the worst case for this resonance. The gap is inflated by an unknown amount. Which is exactly why the dose-response, not the curve, was the pre-registered criterion: it compares the quad against itself at different RPMs within one flight, so it does not care about the weather.

Slope across the band, which is the number that matters:

Sitting in the resonance window has stopped mattering. The criterion passed. Two more things beat the foam flight on the same log: post-filter roll noise 0.34 °/s at 41.2 dB attenuation, the cleanest of the session against 0.67 °/s and 31.8 dB for the foam; and the flattest motor balance I have recorded on this quad — deviations of −0.1 / −4.2 / +2.5 / +1.7 percent, a 6.7-point spread where every earlier flight ran 17–25, front/rear split +1.7%, zero clipping.

What that log cannot tell me: it was indoors and it sampled one RPM bin, 80 of the 84 clean seconds at 300–325 Hz. I had asked myself for 3–4 slow throttle sweeps and flew a hover instead, so there is no structural curve and I cannot locate the mode frequency from one RPM slice. The raw pre-filter number also looks worse than the foam — 39.1 °/s against 26.0 — but the foam flight was outdoors in 4.71 °/s of wind and this one indoors at 1.99, and still air is the worst case, so that comparison loads the dice against the calm one. The only like-for-like number I have is indoor against indoor: the pre-foam, pre-rotation indoor flight read 54 °/s at 300–325 Hz and this one reads 39, about 28% better. Real, but one bin.

Three: outdoors, where the trade-off flipped

121 s clean, outdoors, 5.51 °/s of wind, zero config changes, and finally proper RPM coverage: 8 of 12 bins at 4 s or more against 5 for every previous flight. Best dataset of the exercise.

The pre-registered criterion held, now outdoors:

The structure-fixed feature agrees: 363 Hz with the TPU, 368 Hz with the foam, 255 Hz with neither. Both stiff solutions land in the same place — stiffening moved that feature up by about 110 Hz and it stayed moved.

But the foam is still the quieter mount. Outdoor against outdoor at matched prop RPM, the fair comparison I had been waiting two days for:

Mean across reliable bins: 26.2 °/s for the foam, 33.0 for the TPU, about 26% worse. The curve is less flat too — flatness 1.13 for the foam, 2.14 for the TPU, worse even than the 1.85 of no mount treatment at all — with a peak back at the low end, 44 °/s at 275–300 Hz falling to 21 by 425.

So the amplification mechanism is dead, but the overall vibration level is up. Different statements, both true.

And then the camera got jello back

The part I did not predict, and the thesis of this post arriving from the opposite direction. Energy in the 250–450 Hz band, which is what a rolling shutter turns into jello:

Low-frequency shake is now largely imperceptible in the air, and jello is back on the footage.

And my first explanation was wrong. I claimed removing the VTX grommets had "rigidly bonded the camera to the canopy." It had not: the VTX is a bare board and the camera is mounted on the canopy, not on the VTX. Those grommets were suspending a board with nothing on it — dead weight and one more suspended mass free to resonate. Removing them was not the jello mechanism.

The mechanism is the canopy-to-frame path, because that is what the camera rides on. TPU in those gummies stiffened it, and a stiffer path transmits more frame vibration straight into the camera. Unlike foam, filament adds stiffness without meaningful damping — it couples without absorbing. So jello risk is a product, not a level:

jello ≈ (vibration at the frame) × (transmissibility of the canopy mount at those frequencies)

The FC gyro measures only the first term. Nothing in a blackbox log measures the second, which has consequences for one of my own tables further down.

One caveat about the measurement itself: stiffening the gyro's own mount changes what the gyro reports, not only what the airframe does. A rigidly mounted gyro couples more faithfully to the frame's real motion, so part of the increase in these pre-filter numbers is better coupling to the truth rather than a worse airframe. I cannot separate those two using a gyro that is itself part of the experiment.

Four and five: all of them on one chart, and the small pad

The obvious next move was to put the VTX grommets back and keep the TPU in the gummies — different paths, different symptoms, no reason to sacrifice the camera's isolator to stiffen the FC mount. Removing just the front TPU, the one gummy joining canopy to frame at the nose, pulled the structure-fixed feature from 363 Hz back down to 280 Hz and nearly halved its dominance. One gummy. That is how localised this turned out to be.

Every curve outdoor, binned by mean prop frequency, only bins with 4 s or more of dwell:

Then the configuration I settled on, jello being the deciding factor: all TPU removed, original gummies back, and a small foam pad glued near the connector so it damps without sitting over the hot side of the board:

Original gummies back in, plus one small foam pad near the connector. Note the size: the pad that actually killed the resonance was much larger and sat between the boards.

Note the size. The foam that damped the resonance essentially away was a large piece between the boards, considerably bigger than this. The small pad was a deliberate compromise: enough damping to be worth having, small enough that the ESC side keeps breathing.

It was the worst configuration of the five, and worth being blunt about how badly:

Highest mean vibration of any configuration — worse than doing nothing at all — with the amplification right back at +66%. And the last column stands out: the structure-fixed mode is 81× above background where every other configuration sits between 4.4× and 8.2×. An order of magnitude sharper. Pitch says the same at 9.4 °/s pre-filter, the worst of the five.

A small pad does not damp a mode, it adds a lightly-damped spring in one spot. The large pad worked because it was large enough to absorb across the whole interface. And that sharp 311 Hz mode explains the occasional jello even with soft stock gummies back in: isolation is not absolute, and a mode this dominant pushes through a soft mount some of the time. Which is exactly what I saw — not constant jello, jello sometimes.

Verdict: the small pad comes out. The honest options are the large pad, which measurably worked and cooks the ESC side, or bare stock gummies, which never gave jello but leave the mode running in the airframe. There is no single dial: stiff transmits vibration to the camera, soft leaves the mode free to run, and the only thing that fixed both at once was damping.

Three scares that were not what they looked like

Late in the front-TPU flight I pulled a split-S and the quad snapped as if it had hit something. There was nothing to hit, and the log agrees: peak 3.8 G against 9.8 G for a known floor bump and 9.6 G for a known crash in the same session. Not the radio either — rxSignalReceived and rxFlightChannelsValid never dropped, failsafePhase stayed 0 all flight, and the RSSI minimum sits at t≈39 s, nowhere near the event.

What actually happened, at t = 86.2–86.5 s:

185.95  motor2 driven to the floor (248 -> 128), its RPM falls 6450 -> 2700
286.20  yaw I-term saturates at -230 and pins there
386.20  motors 3 and 4 hit the 2047 ceiling WHILE motor2 sits at 128
486.40  gyro roll -637, pitch -295, yaw +278 deg/s ... commanded yaw = 0
586.45  yaw reaches 346 deg/s, entirely uncommanded

Across 85.5–87.0 s, 17.6% of frames had a motor at the ceiling and 30.4% had a motor at the floor. The mixer ran out of range at both ends simultaneously, leaving no differential authority to answer the sticks. Commanded yaw p99 was 19 °/s; the quad delivered 370. Cause: a high-throttle split-S on a sagging 1S pack with the yaw I-term already pinned fighting the standing yaw bias measured earlier. Thrust and authority running out at the same moment.

The loose-connector theory, tested. My first instinct was a battery connection briefly dropping out. The log says no:

• Fitting Vbat = V0 − I·R over the flight gives about 35 mΩ, the healthy end of normal for a 1S pack plus wiring.
• Zero frames show a voltage deficit unexplained by current draw — exactly the signature a connector letting go would produce, and it is absent.
• The RPM collapse hit one motor, not four: at the worst instant motor 2 was at 2600 RPM while motors 1, 3 and 4 were at 21 417, 14 617 and 23 033. A pack disconnect starves all four.
• At that instant motor 2's commanded output was 238 of 2047. The mixer put it there. It was not power-starved, it was told to stop.

Two caveats: the current sensor scale on this board is unverified, so treat the 35 mΩ as indicative, and the regression cannot separate load sag from pack depletion over the flight, which is why its R² is only 0.28 — the absence of any sharp unexplained step is robust regardless. The instinct that low RPM is dangerous is right, though: 2600 RPM is low enough to risk desync on spool-up. It just did not get the chance. dyn_idle held, with only 0.04% of airborne time under its 3000 RPM target and the longest continuous excursion lasting 4 ms.

Then two more jerks, and it was not the tune. One turning, one diving, plus a third at the very end. I checked the tune first because that was my own suspicion: the configuration is byte-identical to the previous flight. Radio clear again — no dropped frames, failsafePhase 0, RSSI minimum 329 with nothing near either event.

The zoom on the first is unambiguous: motor 2 driven down 293 → 146 → 124 and pinned near 150 for roughly 400 ms while motor 4 rides the 2027 ceiling. Pack sags 3.81 → 3.51 V. Yaw walks out to 86 °/s against a roll command of 47 and a yaw command of nothing. Then motor 2 spools back — 433, 562, 735, 917 — and it flies again. Same failure as the split-S. Flight-wide, 2.74% of frames have a motor at the ceiling, and motors 3 and 4 do almost all of it (1.60% and 1.30%).

One tune-adjacent finding is real: yaw I-term swings between −255 and +271, railing in both directions. That is the standing yaw bias consuming authority before a maneuver even starts, and fixing the imbalance frees more headroom than any gain change will. No dyn_idle problem here either — 0.076% of airborne time below the 3000 RPM target, longest excursion 4 ms.

And an impact I did not report at the time. At t = 109.83 s there is a 12.9 G spike with pitch hitting 2000 °/s, and the log ends. The known floor bump was 9.8 G, the known crash 9.6 G; this was harder than both. Worth a close look at the frame and props, whatever it was.

Everything I got wrong

The retractions are the most useful content in this post, so here they are in one place. Almost every one is a real measurement, competently executed, aimed at the wrong quantity — the failure mode I now watch for hardest.

"Stiffness, not mass" was a false dichotomy

I first wrote the foam result up as "stiffness, not mass," backed by a hover-RPM mass check (−0.8%), a mode shift from ~325 Hz to ~395 Hz, and a confident "≈48% stiffer." All three parts were wrong or unjustified. Coupling two previously-independent bodies changes effective stiffness, modal mass and damping simultaneously, and there is no way to separate them from this data. I framed a question the experiment could not answer and then answered it anyway.

The hover-RPM mass test answered the wrong question

Hover RPM measures total AUW. Coupling the canopy does not change total AUW, it changes modal mass — the mass participating in that particular mode. Using one to dismiss the other is a category error, and it is the mistake I am least happy about, because it is the kind that feels like rigour while you are making it.

What I can show is a controlled comparison: light pack versus heavy pack, foam absent in both, only the pack swapped.

Added sprung mass moved the structural feature down while the forcing went up — √(k/m) behaving itself. The coupling model, that tying the canopy to the frame removes a relative degree of freedom rather than merely shifting a spring constant, is at least as well supported as the stiffness framing and better supported on the mass side.

Practical consequence: gummy balls couple FC to frame, the foam coupled canopy to FC and frame. Harder balls alone would not reproduce that, which is why the next experiment stiffened the gummies from the inside rather than swapping durometer.

The 325 → 395 Hz shift and the 48% figure are withdrawn

Two implementations of the same "structure-fixed frequency" detector disagreed badly on identical data: one said 322–329 Hz at 120× dominance, the other 255 Hz at 6×. The cause is visible once you look — with four motors spread ~30 Hz apart, a 40 Hz RPM slice gets contaminated by whichever motor is slowest, so "mean RPM" is a poor label for what is in the bin. Both numbers are gone.

The amplitude results survive method-independently; they do not depend on locating the mode at all. The foam produced a large, real reduction — that was never in question.

A metric that lied to me for several rounds

For several rounds I scored the wind-shake verdict as a single global ratio, shake / wind, and got 0.777 → 0.798 → 0.791 → 0.754. Read as "−4.4%, within noise, no real improvement," and I nearly wrote off the foam on that basis.

Artifact. Shake versus wind is not proportional, so a global ratio depends entirely on where in the wind range you happened to sample. Bin by instantaneous wind level instead and compare only the bins both flights actually sampled:

1heavy no foam -> +foam : 6.21 -> 5.29  = -14.8%   (5 shared bins)
2original      -> +foam : 6.65 -> 5.29  = -20.4%   (5 shared bins)

About 15% less wind shake at matched wind, not 4%. And look at the shape: all four flights agree in the lowest wind bin (2.27–2.56) and diverge only as wind rises. That agreement at the bottom is the signature of a calibrated measurement — the flights are not offset from each other, they have genuinely different slopes.

Audited at the same time, which explains a lot of earlier flailing: every flight up to that point reached ≥4 s dwell in only 5 to 7 of 12 RPM bins.

The column I called the "jello band"

An earlier version of this post used the 250–450 Hz column to rank which mount gave the worst jello. That was wrong, and it inverted reality. That number comes from the gyro on the flight controller — it describes what the FC experiences. The camera is on the canopy behind a separate mount, so FC vibration only becomes jello to the extent the canopy path transmits it.

What actually happened, observed on the footage rather than inferred from a gyro:

Entirely consistent, and the opposite of what my table implied. With soft original gummies the frame can shake hard — 34.5 in that column — and the camera never sees it, because the canopy mount is not transmitting. Stiffen that mount and the same vibration arrives at the sensor. Soft gummies are the best case for jello, not the worst. The footage was the only valid instrument here and I should have deferred to it sooner.

A bug in my own analyser

My first step-response report proudly announced "overshoot 0.0%" on all three axes. Exactly zero, all three. The function normalised each response by its peak, which pins overshoot at zero every time, by construction. Fixed to normalise by steady state. If a metric comes out suspiciously clean on every axis at once, the metric is broken.

The confound I should have flagged much earlier

This is not a Meteor75 Pro II. It is a Pro II shell with Pro internals, and that includes the motors: I kept the original 22,000 KV units, where a stock Pro II ships 21,000 KV.

Being precise, because this is easy to overclaim. It does not move the hover excitation frequency — hover RPM is set by the thrust the airframe needs, not by KV, so a lower-KV motor hovers at the same RPM at a slightly higher throttle position and the 325–365 Hz window is not a KV artifact. It does change RPM-per-throttle, torque constant and current draw, so stick-to-RPM mapping and the saturation headroom behind that split-S departure are both affected.

The larger point stands: a stock Pro II with the O4 Wide differs from this build in motors, canopy loading and mass distribution at once. Every mount conclusion here is measured on a hybrid and I cannot claim it transfers to a factory unit.

And the big one: I spent a week measuring the wrong band

Now the correction that is more useful than any individual result above.

I spent a week characterising a structural resonance at 320–345 Hz, and I measured it well. Prop changes, canopy rotation, foam, TPU in the gummies, five mount configurations, dose-response curves, mode frequencies, transmissibility. All real, all repeatable.

None of it was the thing I had actually set out to fix.

What I kept writing in my own notes was this: the quad shakes, sometimes, for a long time, flying straight as well as turning. If it were a resonance it would do it all the time. That objection was mine, it was correct, and I talked past it for days because the resonance was so much more fun to measure.

So I went and looked at where the uncommanded motion actually lives — uncommanded meaning gyro minus setpoint, what the quad does that I did not ask for:

Ninety-seven percent of it is below 25 Hz. My entire analysis lived at 80–780 Hz. I was measuring the band my tools were sharpest in, not the band the symptom was in.

What it looks like at the exact frame I marked

Steady fast flight, throttle 1568, props at 447 Hz, sticks essentially still:

Seventy-seven times more motion than the stick asked for, and not one saturated frame.

Two things that rule out my earlier explanations:

• The filters are not doing this. Band-limited 14–23 Hz, unfiltered gyro reads 6.38 °/s and filtered reads 6.37 — a ratio of 1.00. The filters neither create nor remove it, because that band is deliberately passed through so the loop can control the aircraft. Filtering is irrelevant here, which is why every filter change I made left it untouched.
• It is not the 320–345 Hz mode. It does not track RPM the way a prop-order forcing does, its Q is far too low for the structural feature I had been measuring, and its apparent peak wanders between 10 and 30 Hz across RPM bins with weak prominence.

It is at least two different problems, split by regime

That is the part I got wrong by trying to force one story:

The jerks and the sustained shake are related but not identical: the jerk is the saturation case at its worst, the everyday shake in fast flight is a rejection problem with the mixer nowhere near its limits.

A resonance is satisfying to chase — it has a frequency, it responds to mechanical changes, it makes clean charts, and every intervention produces a measurable delta, so it feels like progress. It took me writing "if it were resonance it would shake all the time" in my own notes three separate times before I stopped defending the frame I had built and went to look at the raw uncommanded motion. The instrument I trusted most, the blackbox gyro spectrum, is why I got stuck. It is excellent above 80 Hz and I read it constantly. The answer was underneath it the whole time.

Why the foam helped after all

That leaves one thing unexplained: if the shake is a low-frequency authority problem, why did a lump of foam between two boards help so much? Foam adds no thrust and no mixer range. It should be irrelevant.

It is not. Matched conditions — steady flight, throttle 1380–1560, impacts excluded, so aggression differences are not driving it:

corr(vibration, 1–8 Hz uncommanded) = +0.92. corr(vibration, headroom) = −0.92. The large foam wins on both counts, the small pad loses on both with 2.2× more low-frequency wobble than the large pad. So the mount does reach the thing I actually care about — just not the way I assumed.

Not via motor jitter, which was my first guess: vibration into the D term, motors jitter, jitter eats mixer range. Measured, it does not carry the weight. Motor jitter runs 5.3–7.1 counts RMS, roughly 1.6–2.1% of the available range — real, and far too small to explain a loss of authority.

The explanation that fits: the O4 and the canopy are a significant mass, and on soft gummies that mass can move relative to the frame. That makes the aircraft a two-body system — the loop commands the frame, and the canopy follows late and overshoots. That relative motion is uncontrolled by definition. No PID gain reaches it, because the gyro is on the other body. And it shows up exactly where the symptom lives: slow, uncommanded, 1–8 Hz wobble I can see in the goggles and the tune cannot touch.

Foam does not stiffen that interface, it damps it. Damped, the two masses move as one and the loop is finally controlling the whole aircraft rather than one body bolted to a swinging one. Which retro-fits the whole sequence of experiments:

• Large foam best — it damps the relative mode across the whole interface
• TPU worse than foam — stiffness without damping still permits resonant exchange, it only moves the frequency
• Small pad worst — too small to damp anything, and it introduced a razor-sharp mode at 81× dominance where every other configuration sat between 4.4× and 8.2×
• No tune change ever helped — because it was never a gain problem

So the mount investigation was not a rabbit after all. For a week I was reporting the right intervention with the wrong reason attached.

Honest limits. Five configurations across two days with different packs, weather and aggression. An r of +0.92 on five confounded points is suggestive, not proven. And the within-flight correlation between vibration and shake is near zero or slightly negative (−0.02 to −0.27), which says this is a property of the configuration rather than moment-to-moment cause and effect — consistent with a structural-dynamics explanation, inconsistent with a noise one.

The shake I could not see, and the term that was driving it

Everything above measured 80–780 Hz. But I could see the shake with my own eyes, and nobody can see 350 Hz — coupled to the camera it shows up as jello, a different symptom entirely. I was right about what I was watching and wrong about where to look for it, and the number was in my own band table the whole time: 58.6% of the uncommanded motion sits in 1–4 Hz. So I band-limited the original log to 0.5–3 Hz at the exact frame I had marked, and it is there:

One-point-one hertz on roll: a full cycle in under a second, gyro +21.9 °/s at t=45.11 swinging to −21.2 °/s at t=45.91 while the stick asks for almost nothing. That is a wave you can watch with your eyes, and exactly the thing I had been describing since the first flight.

Which term is doing it

Decomposing the oscillating roll command inside 0.5–3 Hz:

The integrator is 93% of it, and larger than the sum, which means P is partly cancelling it rather than working with it. The phase is the proof:

1I-term phase vs gyro:   roll +105 deg    pitch +90 deg
2I-term own period:      roll 0.97 s      pitch 0.83 s
3shake period:           roll 0.90 s      pitch 0.56 s

A healthy integrator opposes the error, near 180°. This one sits at +90 to +105° — quadrature, in step with the error's rate rather than against the error itself. An integrator in quadrature has stopped correcting and started pumping: it winds up while the error is still growing, peaks after the error has already reversed, and hands energy to the next half-cycle. Its own period matches the shake period on both axes.

The motors confirm it is being commanded rather than suffered: roll differential 123.4 counts RMS inside 0.5–3 Hz, coherence 0.84 with the gyro.

Why this fits what I was actually seeing, and my earlier stories did not

• I could see it and could not measure it — 1 Hz is visible airframe motion; 350 Hz is jello. Two symptoms, and I conflated them for days.
• Fresh pack, shaking immediately — integrator dynamics do not care about voltage, which killed my sag-compensation theory outright.
• Straight flight and turns alike — the I term winds up against any sustained error, including the standing imbalance measured earlier in this post.
• Sustained for seconds — a limit cycle persists once it locks.
• No mechanical change ever fixed it — it lives in the controller, not the airframe.
• The foam still appeared to help — damping the plant adds phase margin, which raises the threshold for the cycle to sustain. The mount was never the cause, but it did move the margin.

And the gains line up with the axes: i_roll is 80 against stock's 67, i_yaw the same +19% — the two axes that oscillate. Pitch sits below stock and oscillates roughly a third as hard (2.88 versus 8.31).

Where it stands, and the test that settles it

The small pad comes out. Beyond that, in order:

1. Go back to the large foam pad. Best measured result on both metrics — the 80–780 Hz vibration and the 1–8 Hz uncommanded motion — and it never gave jello.
2. Solve the heat differently instead of shrinking the damper. A large thin damping layer spanning the interface — viscoelastic tape rather than a thick block plugging the gap — keeps the damped area without being an insulating bung. Or keep the large pad and route it clear of the ESC FETs, which are the hot part, not the MCU.
3. Stop chasing stiffness. Damping beats stiffness here, measured rather than assumed.
4. Secure the O4 to the frame, not only to the canopy. Less independently-moving mass is the root fix; everything else is mitigation.
5. Get hold of a stock Pro II with the Wide unit. Every conclusion here is measured on a hybrid, and that is the honest way to find out what transfers.

And then the one test that settles the diagnosis, which is a CLI paste rather than a screwdriver:

1profile 0
2set i_roll = 50
3set i_yaw = 50
4save

Deliberately well below the factory 67, because I want an unambiguous answer rather than a subtle one. If the 1 Hz wave collapses the diagnosis holds and I walk I back up to find the ceiling. If it survives at 50, the integrator is innocent and I am wrong again.

Worth stating plainly: this is the third mechanism I have proposed for the same symptom. The resonance chase was measuring the wrong band; the saturation story explains the violent jerks but not this, since at the marker there is not one saturated frame. Both are still in this post, wrong parts included, because the sequence is the honest record of how the diagnosis actually went.

Method notes worth keeping

How to read the numbers first. Figures in this post come from two different analysis windows, and they are not interchangeable. Most tables use outdoor flight binned by mean prop frequency, counting only bins with 4 s or more of dwell — that is the set that reads 37.7 / 26.2 / 33.0 / 30.1 / 41.0 °/s across the five mounts. The vibration-versus-wobble comparison instead uses steady flight at throttle 1380–1560 with impacts excluded, which is the set that reads 38.3 / 26.0 / 31.0 / 26.6 / 42.5. Same quantity, different slice, so the two sets disagree by a few percent and neither is wrong. Where two figures for the same thing differ slightly — 34.8 against 34.5 in the 250–450 Hz column, +66% against +65% for the untreated slope, 0.789% against 0.812% for m1 clipping, 0.04% against 0.076% under the dyn_idle target — the difference is the impact-exclusion pad or the log being different, not a correction. Post-filter roll noise appears as both 0.50 and 0.67 °/s for the foam because one is a Welch integration above 60 Hz and the other an 80–780 Hz bandpass RMS; they are two metrics, not two answers.

Practices that repeatedly changed the conclusion — not general advice, things that actually flipped an answer in this specific week:

flowchart LR
    A["Raw .BBL"] --> B["Decode EVERY log in the file<br/>including the truncated last one"]
    B --> C["Exclude crashes and bumps<br/>trim ~15 s after arm / before disarm"]
    C --> D["Establish the noise floor FIRST<br/>CV 9% -> +/-10% is nothing"]
    D --> E["Compare at MATCHED RPM<br/>the resonance makes raw comparison meaningless"]
    E --> F["Check dwell time per bin<br/>4 s through a resonance != 70 s parked on it"]
    F --> G{"Is the number bigger<br/>than the noise floor?"}
    G -->|"no"| H["Not a result.<br/>Say so."]
    G -->|"yes"| I["Validate settings against data<br/>motor_poles was a SETTING"]
    I --> J["Never assume gyro units<br/>auto-detect and report"]
    J --> K["Report it — with the confound"]

    style D fill:#244d68,color:#fff
    style H fill:#915d52,color:#fff
    style K fill:#244d68,color:#fff

• Decode every log in the file, including the last one even if truncated. Battery pulls and crashes routinely truncate the final log, and that is often the interesting one.
• Exclude crashes and bumps and trim ~15 s after arm and before disarm, before drawing any conclusion at all.
• Establish the noise floor first. CV 9% meant several "improvements" were nothing.
• Compare at matched RPM, always, and watch dwell time — a 4 s excursion through a resonance cannot build the amplitude 70 s parked on it does, so thin bins mislead you in a direction that looks like a result.
• A logged value can be a setting, not a measurement. motor_poles was validated against the data rather than trusted. And never assume gyro units — auto-detect and report.
• Check which band the symptom is in before characterising a band. This is the one that cost me the week.

Where I leave it, and what is next

The foam is out. The TPU inserts are out. What is in instead is a little B7000 inside the grommets — enough to stiffen them a bit without going back to a filament shim or a thermal blanket over the ESC side. That is the whole mechanical change, and I am going to fly it in a range of weather before I touch anything else.

And I have decided to stop fighting one of the two symptoms.

I accept the shake in wind. Not because it is solved — it is not, and the 1 Hz integrator question is still open — but because it is the recoverable one. Shake I can stabilise in DaVinci. Jello I cannot. There is no filter, no Gyroflow pass and no RockSteady setting that un-bakes rolling-shutter distortion out of a frame, so between the two, the one I refuse to ship is jello.

That trade is not theoretical. I flew a project shoot recently where people were playing volleyball and I orbited around them. The previous canopy setup shook violently too — and it produced so much jello that I came away with no usable footage at all. Nothing to grade, nothing to stabilise, nothing to keep. This floaty canopy will shake in the same conditions, but it will hand me frames that are sharp enough to fix. Wobbly and recoverable beats steady and ruined.

There is also a simpler answer I should admit to: if I want cinematic footage of people playing sport, I should be flying the Pavo20, not a 75 mm whoop. Snake weighs almost nothing, and almost nothing is exactly what wind pushes around. Choosing the right airframe would have saved me most of this week.

The open list

• Fly the i_roll / i_yaw = 50 test and log it. One CLI paste settles whether the 1.1 Hz wave is the integrator or something else, and I have been putting it off in favour of mechanical changes that were never going to reach it.
• Log a 60 s flight with real stick movement. Flick-and-hold, hard reversals, a few punches. Everything I know about this quad's step response rests on six usable steps from a log that ended in a crash.
• Watch the B7000 grommets over a few packs — does the stiffening hold, and does jello stay away as the glue cures and settles.
• Try a stock Meteor75 Pro II with the O4 Wide. Every conclusion here is measured on a hybrid: Pro internals, Pro motors at 22,000 KV, in a Pro II shell. I cannot claim any of it transfers to a factory unit, and a stock airframe would tell me which half of this was the build and which half was me.

I will post the numbers either way, including the ones that make me look silly. There have been plenty of those already.

What I actually bought

The thing I set out to fix was jello, and I fixed it — by buying a frame whose canopy holds the camera away from the shaking. The thing I did not expect to buy along with it was a soft spring between the flight controller and the airframe, tuned by accident to a frequency four motors pass through every time the wind pushes the quad sideways.

Better isolation gave me footage Gyroflow can rescue on a bright day and a gyro trace that looks like an emergency. Those are the same change. And the shake I could see the whole time — the one that started this — was never in that trace at all. It was a hertz and a bit, underneath everything I was measuring, driven by an integrator winding up against an imbalance I had also measured and filed as unrelated.

A week of logs, several retractions and one very embarrassing analyser bug later, the only lever that moved the structural problem was a piece of foam. It works by damping — that part I can now say. Whether the 1 Hz wave belongs to the same story or to the integrator alone is one CLI paste away, and I will post the numbers either way.

Craft: Snake — Meteor75 Pro II frame and canopy, Matrix 1S 3-in-1 FC, narrow-FOV DJI O4. Betaflight 4.5.1, 3.2 kHz loop, blackbox at 1582 Hz. All figures measured from blackbox logs; clean sections only, crashes and bumps excluded. 169 flights / 15,574 s logged at time of analysis.