Tuning the LQG flight mode (bis)

This site is best viewed in a modern browser with JavaScript enabled.

Tuning the LQG flight mode (bis)

raemin_bis

This has been somehow requested in the previous post by fxbisto, but as I've just managed to read again my log, I can post here a few questions with details. Basically the quad flights quite well in axis lock mode, but when switching to LQG it tends to overshoot when exiting corners.

See 1:59 in the video (note that 1:59 flight time is 0:49s on the video abstract). That's a fatshark video and they only record even frames, so it is not as obvious as in the live video feed, but the 1s "glitches" are still visible.

And here is the corresponding log, as can be seen, the actuator curves do have quite a few peaks in LQG compared to axislock.

logvieweranalysis.png

It used to be horrendous with motor curve fit set to 0.85, now I am down to 0.69, and only have the issue while cornering. Just wondering if reducing Q2 would help?

By the way I am very happy to have my logs back! (had to downgrade the scipts and use "Wired" branch)

raemin_bis

A bit later during the same flight, similar glitches:

capturelqg.png

glowtape

Those wiggles were something ufo and I were trying to track down. Seemed to happen mostly on light and/or overpowered quads.

Playing with Q2 on either the RTKF or LQR is worthwhile. Either the torque is modelled not well enough (RTKF), or its going gung-ho on the motors via the torque term (LQR).

I have a bunch of other things on the plate with the bits of time spending on dR, but when I can finally be assed to do something about realtime ESC telemetry, that could help (since we have data sort of looking like actual torque).

raemin_bis

That's a light quad. I presume the problem is that autotune/hovering is at 10% throttle. Obviously at 75% throttle that's another beast. I've tried to make the best motor curve fit based on MiniQuadTestBench dyno tests, but even-though it's hard to extrapolate from data collected at such low throttle.

motorcurvefit.png

Assuming beta is underestimated during autotune, should I slightly reduce LQR Q2 (or increase it?)

Before the gyro starts to shake one can see a 30% actuator command.

Can I assume that reducing LQG bias to 25% will reduce the initial 30% actuator peak?

capturelqgbis.png

My wiggles were really visible on the first -turn but looking at the log and video it seems that on the following turn there is the same big initial actuator peak, but less wiggles and RTKF looks different. Actually there are only 17 RTKF logs between 117.035 and 120.655 , so the RTKF is more or less flat and actuator data shakes like hell, second pass there are approx 265 logs so the RTKF curve matches the gyro data and actuators are managed more smoothly.

Can i assume RTKF does not always operate, but when it does it seems to properly handle my actuator? In other words no point in messing RTKF Q2?

glowtape

Yea, ufo's quads also had such low throttle. I presume it's a curve fitting/torque issue.

As far as what to do with Q2, I'm not entirely sure. Back then, he kept sending me logs and I tried making suggestions. Without my own superlight quad, it's sort of hard. If you tune it, scale it in percentages, don't go hogwild. I.e. add percentages on the existing values.

The RTKF runs in lockstep with stabilization. Seems like data dropped during logging.

--edit: It's possible beta on roll and pitch isn't entirely right. Maybe it guessed too low for such a powerful quad. But if you start tweaking that, do note that it's an exponent to e. Increasing beta by one implies the quad were 2.71x more powerful.

raemin_bis

Thanks for the hint(s). Long story short: had to increase LQR Q2 a lot. Very pleased with how it feels now.

First It's not obvious, but one has to INCREASE Q1 Q2 in order to reduce actuator effort. Abstract from the code:

Q1 and Q2 in the state weighting matrix penalizes rate and torque.

R seems to work the other way round (decrease R in order to reduce actuation effort). Not sure though...

So here is my original log with default settings. On this test I made a few pitch-up commands with the auto-leveling that was fighting back to take the aircraft back to level. Only displaying actuator command and gyro.

log-0.png

Attempting to play with Bias Limit, Q1 and Rate (non satisfactory, and even a sluggish feeling after modifying R)

log-4.png

log-5.png

log-6.png

Now the graphs for Q2

log-1.png

log-2.png

log-3.png

next graphs do use another layout (sorry for the mismatch gyro curve is not purple anymore)

log-7.png

log-8.png

glowtape

FYI you're penalising the error. If you raise Q1 or Q2, it'll try harder to drive the rate and torque errors, respectively, to zero. I think we chose the current values based on noise and such. I'll countercheck your new ones next time I go flying. Interesting find.

Are your Q2 values in the picture multiplicators or absolutes?

raemin_bis

I still have to find a way to "dampen" it but this is working much better now with
1) BLHeli adjustments (reduced startup power),
2) motor curve fit adjustment,
3) autotune adjustment (actuator effort raised to 0.25),
4) increase LQR Q2.

The Q2 values in my graphs are what I've put on the system tab. So absolute values. Modifying Q2 has cleared the low throttle errors I had, making banking extremely locked. Also made a few blheli adjustments (reduce startup power) and raised motor curve fit from 0.68 to 0.72) in order to reduce low throttle audible clicks. With all these adjustments low throttle is reasonably good.

From then on curing high throttle oscillation was another story. I think autotune is not accurately analysing the aircraft at 11% hover throttle, so I identified which actuator effort was causing problems (my logs do show that a 25% actuator peak produces many subsequent glitches) so decided that this would be what to use for autotune actuator effort (0.25). Autotune on this basis gave a slower tau and higher beta.

I've tried these settings in strong gust of winds (45kmh) and it flies much better than in acro. There is still a bit of overshoot: looking a the log the actuators are oscillating a bit but there are no visible effect on the video footage.

Only issue left: my older batteries (7mOhm per cell) are sagging, only my brand new DemonRC lipo could sustain the ampdraw. So ideally I would need to dampen it (or only use fresh batteries!).

glowtape

Hmmm I have to reinstall MATLAB and throw these Q2s against the theoretical model. AFAIK high Q2 values ought to have a huge dampening effect, especially that high, so I'm surprised it actually works. How does ActuatorCommand look, anyway?

raemin_bis

Well on this post I was increasing LQRegulator Q2 untill the gyro curve looked correct. The figures are displayed without the exponent so (Q2=1.2 in my graphs means 0.00012 in the system tab).

Based on peabody's blog it seems Q2 and Q3 from RTKF do increase actuator damping. So as a next step I have increased RTKF Q2&Q3 until the actuator response is not overshooting so much. On the following graphs I increased Q2&Q3 by a 10% to 80% factor. There is no significant difference on the gyro side, but It seems that at +70% the actuator response is more consistent, at +80% the gyro curves starts to suffer.

log001std.png

log002damp11.png

log004damp13.png

log005damp15.png

log008damp17biaslim25.png

log009damp18biaslim25.png

All in all with +70% on RTKF Q2/Q3, the FC has to send 4 consecutive corrections in the same directions in order to stabilize the aircraft. Ideally I would prefer one or two burst, so maybe I've dampen it too much. At least it is not issuing order in opposite direction as was the case before, so hopefully the batteries will like it better.

glowtape

The RTKF is the Kalman filter that essentially models the torque and the bias terms, latter which is essentially an integral. Raising Q2 will use more of the sensor value to correct the model guess, but how far that's valid is anyone's guess, since rate is a linear term and torque a derivative based on actuator range. Raising Q3 speeds up how fast the bias/integral rises and falls, too much isn't always a good thing.

That's why eventually, having actual torque values would be interesting, i.e. taking realtime telemetry from the ESCs, remapping it to actuator range (probably based on hover throttle and some curve mapping), and then do an inverse mix to get roll, pitch and yaw torques.

Also, raising Q2 will introduce noise intro the torque term, if the gyro is noisy, which gets then amplified by the LQR.

But anyway, as you see, there's tons of parameters to mess around with, so it's a surprise it works largely with the average powered quads. It was quite a pain in the ass to get a good set to begin with

--edit: That ratcheting you see in ActuatorDesired, specifically the second to last image, is the LQR trying too hard. The interval should match the command rate of your remote. That's why the Q1 and Q2 values were that low as they're shipping as default. If you raise Q1/Q2 or drop R, this will get worse.

--edit: You could also try eliminating the ratcheting another way. We have control smoothing implemented. It's ManualControlSmoothing in the StabilizationSettings UAVO. Normal blunts the edges using some predictive action, whereas Linear is plain linear interpolation. The duty cycle controls over what period of the control interval for both options. I should probably finally expose that option in the UI.

raemin_bis

Thanks GlowTape , for spending all this time explaining how to tune my beast! Took me 64 flights logs in order to understand how the parameters are working...

Before

logs-initialsettings.png

After

log030-rtkf-q2-35ok.png

And these are roughly speaking the parameters that had the most impact

logs-initialsetting-comments.png

LQR-Q1=0,00001200
LQR-Q2=0,00012500
RTKF-Q2=0,00000350
RTKF-Q3=0,00000130

The two major differences compared to stock settings are LQR-Q1 (much reduced) and RTKF-Q3. The other settings do only have a marginal effect. Tuning RTKF-Q2 is tricky: too low and the actuator effort bounced up and down in opposite direction, too high and there is some ratcheting on the actuator effort (similar to the "second to last" screenshot of my previous post, as quoted by glowtape).

At the end of the day I could not totally get rid of the little peak on the gyro graph, but at least it is much more smooth and the feeling is amazing!

fxbisto

Hey guys, I've had to drop away from RC for a little while due to health issues but hopefully soon I'll be getting back into it. @raemin_bis it's great to see extensive logging data to go with this; I wish I could have joined in with you on this tuning mission!

For me it's always been about the propwash oscillations, and I had started to tame them by changing the Kalman Q values, but that had introduced some vague weirdness in other handling scenarios which I presume meant I needed to penalise the regulator a bit more to get the feeling back.

It's easy to quantify the behaviour I was observing in PID terms as if I were lowering the I-term, however it was initially so aggressive that it kept the P gain low. Reducing that overly-stiff feeling helped lower propwash oscillation but exposed the low effort (or P in this example.)

I was going to play with raising regulator R as it seemed to function as a scale for the entire actuator effort & penalising system. I'm not aware of how it functions mathematically but wondered if raising R and lowering the Q's would increase effort but allow the regulator to be very slightly less strict upon the error.

raemin_bis

If you look at my video we had more or less the same issue: oscillations when ending a turn wich results in backward wind flowwing on the propeller (as glowtape said), reducing LQR Q1 a lot did the trick for me. The issue here is to do it step by step as the response is not linear (as glowtape also said ). So I reduced the settings by 5% incrments untill the issue was cured.

The other parameters had an effect on the graphs, but I would say the "feeling" did not change. I would not have noticed the difference if it were not for the openlog.

By the way, in order to perform my tests, I defined in my radio a -100% roll mixer with 60% offset, and pulled the "trigger" numerous times while in LQG levelling. This is ensures that the oscillations are not generated by the return spring of the stick (as was my problem initially).

Hope this helps.

fxbisto

Thanks for the info. You mentioned the radio mix you setup - I'm assuming that was on a momentary switch? Also did you do all of the tuning in attitude mode or was that just for the step response?

One of the first things I noticed when trying to tune a dRonin quad was the impact of non-smooth setpoint movements on the motors. It's certainly more prominent with higher power-to-weight ratio multirotors, since I found the control smoothing setting it's been permanently set to linear input smoothing @ 50%

I personally don't feel the need for the other (feed-forward?) input filter option, linear is certainly fast enough for my bonkers acro flights. Raising the derivative gamma is a much better way to increase craft response when using the PID controller but definitely requires input smoothing; You're asking D to take more notice of the setpoint when you increase it's value.

In the video there seems to be some noticeable mid-throttle oscillations, did you manage to tune them out or did that disappear when you decreased motor curve fit?

raemin_bis

Yes this was a momentary switch. I recorded the logs in LQL-L mode as this was a way to let the controller perform all the corrections and avoid any noise related to my handling of the sticks.

Mid flight and end of turns oscillations are "cured" for know, but this was more related to LQG settings than the motor curve-fit: I modified motor curve fit but the benefit is marginal below 0.85: once two low this generates some low throttle problems as the dampening/braking command becomes to hard and produces some click-click when hovering.

As a side note, one can roughly say that the LQRegulator makes a prediction and then the Kalman filter sorts of fix the errors reported by the gyro. So it is best to work first on incorrect LQRegulator settings in order to let the filter capture a few errors. On a second phase one can fix the LQRegulator settings.

This being said, tuning the Kalman filters does require logs and graphs, as the effect is quite subtle. By contrast LQRegulator can be fixed visually and has a noticeable effect. So unless you've got an openlager, maybe you should focus on the regulator: when "to high" the sound of the motor is a bit etchy, when "to low" the quad-copter starts to feel sluggish.