Mellinger controller: Attitude control unstable beyond about 30 degrees roll/pitch

[whoenig]

the SE(3) controller from Mellinger/Li should be capable of trajectory tracking with around 60 degrees roll/pitch angle without problems. Our implementation only can get to around 30 degrees. This has been observed by @whoenig and @jpreiss in a trajectory tracking scenario and by @pPatrickK in USC-ACTLab/crazyswarm#71 in a position setpoint step scenario. This might be due to some implementation bug or because of the low thrust-to-weight ratio of Crazyflies.

Transferred from USC-ACTLab/crazyswarm#74.

[jonasdn]

What would be a good way forward for this issue?

How do we determine if this is a bug or if we have reached a limitation of thrust-to-weight ratio?

Any ideas?

[whoenig]

A software verification might be possible with the Python bindings (#602) and simulation of a model with a higher t2w ratio. A hardware verification would require upgraded motors.

[whoenig]

One problem I just found is the omega error: this should be both in body coordinates (equation right above (8) in the paper; Lee uses R^T R_d, see (8) in his paper). I am not sure if this can cause these kind of errors, but it can certainly explain odd behavior for yaw !=0 cases.

[jpreiss]

Did anyone ever do a comparison between the INDI controller and Mellinger for aggressive attitudes? If another controller performs significantly better, that would be evidence that this is a software bug.

[knmcguire]

I've compared the PID and INDI controller, as the INDI controller should be able to withstand sudden distrurbances and windgusts, but not against Mellinger...

Is there a trajectory that forces the 30 deg roll/pitch that we can compare the both with?

[whoenig]

Executing https://raw.githubusercontent.com/USC-ACTLab/crazyswarm/master/ros_ws/src/crazyswarm/scripts/figure8.csv at 0.6 timescale results in 34 deg max roll and 23 deg max pitch. I remember that this results usually in crashes with the mellinger controller.

[knmcguire]

This was very long over due.... but I finally got around to test this. I tested the trajectory on PID, Mellinger and INDI. Although the Crazyflie didn't crash, it did behave more aggressive and wobbly during the maneuver. With bad motors/propellers it probably crashed for sure.

So I'm not sure what is causing it for Mellinger, but perhaps it would be good in general to implement default mellinger gains are a little bit downtuned. Anybody that has those values anywhere?

Videos (might need to downloads to see it)

PID

20220307_140604-converted.mp4

Melinger

20220307_140728-converted.mp4

INDI

20220307_140728-converted.mp4