Simulation and testing problem (inconsistent behavior) about data-driven-dynamics HOT 27 CLOSED

Thank you so much!! I will try flying the drone next weeks and update this comment. I hope that works. Thank you for your all effort))))))

from data-driven-dynamics.

Thank you for reporting this issue. :) We are happy to help you and to further improve the pipeline.
For this we might need some more information. Could you share the full results file you obtain when using the default log? it should be located in the model_results folder.

from data-driven-dynamics.

quadrotor_model.zip
I build the data-driven-dynamics to PX4-Autopilot/build/px4_sitl_default/data-driven-dynamics

from data-driven-dynamics.

Should I open any different log file setting from pixhawk parameters on my quad? Or standart log files are enough?

from data-driven-dynamics.

@AEGEAN99 Could you also share the full log displayed on your terminal?

from data-driven-dynamics.

of course here my termina when training:
`make estimate-model model=quadrotor_model log=resources/quadrotor_model.ulg
python3 Tools/parametric_model/generate_parametric_model.py
--config /home/a3g34n/PX4-Autopilot/build/px4_sitl_default/data-driven-dynamics-master/Tools/parametric_model/configs/quadrotor_model.yaml
--data_selection False
--plot True
resources/quadrotor_model.ulg
Visual Data selection enabled: False
===============================================================================
Data Processing
===============================================================================
Initializing of configuration successful.
Resample frequency: 250.0 Hz
Loading uLog file: resources/quadrotor_model.ulg
Loading topics:
actuator_outputs
vehicle_local_position
vehicle_attitude
vehicle_angular_velocity
sensor_combined
Starting data resampling of topic types: dict_keys(['actuator_outputs', 'vehicle_local_position', 'vehicle_attitude', 'vehicle_angular_velocity', 'sensor_combined'])
Resampling |################################| 6/5
Resampling |################################| 7/5Initializing of configuration successful.
-------------------------------------------------------------------------------
Initialized dataframe with the following columns:
['timestamp', 'u0', 'u1', 'u2', 'u3', 'vx', 'vy', 'vz', 'q0', 'q1', 'q2', 'q3', 'ang_vel_x', 'ang_vel_y', 'ang_vel_z', 'ang_acc_b_x', 'ang_acc_b_y', 'ang_acc_b_z', 'acc_b_x', 'acc_b_y', 'acc_b_z']
Data contains 14775 timestamps.
Computing force features for rotor: front right rotor
Computing moment features for rotor: front right rotor
Computing force features for rotor: back left rotor
Computing moment features for rotor: back left rotor
Computing force features for rotor: front left rotor
Computing moment features for rotor: front left rotor
Computing force features for rotor: back right rotor
Computing moment features for rotor: back right rotor
Cramer-Rao Bounds for force parameters:
vertical_rot_drag_lin 0.012782946245478023
vertical_rot_thrust_lin 0.02641080802919156
vertical_rot_thrust_quad 0.011493031809214254
c_d_fuselage_x 0.013983358667215987
c_d_fuselage_y 0.01915150130118305
c_d_fuselage_z 0.09617539023560966
Cramer-Rao Bounds for moment parameters:
vertical_c_m_leaver_quad 0.6704594437261402
vertical_c_m_leaver_lin 0.43084225232492834
vertical_c_m_drag_z_quad 0.12862700319155235
vertical_c_m_drag_z_lin 0.911563984500324
vertical_c_m_rolling 1.1950003635205881
===============================================================================
Preparing Model Features
===============================================================================
===============================================================================
Initialize Optimizer
QPOptimizer
===============================================================================
Define and solve problem:
min_c (X * c -y)^T * (X * c -y)
s.t. G * c <= h
Initialized with the following coefficients:
['vertical_rot_drag_lin', 'vertical_rot_thrust_lin', 'vertical_rot_thrust_quad', 'vertical_c_m_leaver_quad', 'vertical_c_m_leaver_lin', 'vertical_c_m_drag_z_quad', 'vertical_c_m_drag_z_lin', 'vertical_c_m_rolling', 'c_d_fuselage_x', 'c_d_fuselage_y', 'c_d_fuselage_z']
Fixed Coefficients: Value
-------------------------------------------------------------------------------
Bounded Coefficients: (Min Value, Max Value)
vertical_rot_drag_lin: (0.0, 2.0)
vertical_rot_thrust_lin: (-5.0, 0.0)
vertical_rot_thrust_quad: (0.0, 50.0)
vertical_c_m_leaver_quad: (0.0, 2.0)
vertical_c_m_leaver_lin: (-1.0, 0.0)
vertical_c_m_drag_z_quad: (0.0, 2.0)
vertical_c_m_drag_z_lin: (-1.0, 0.0)
vertical_c_m_rolling: (0.0, 2.0)
c_d_fuselage_x: (0.0, 2.0)
c_d_fuselage_y: (0.0, 2.0)
c_d_fuselage_z: (0.0, 2.0)
[ 1.59066672e-01 -1.38275472e-01 4.19292061e+00 3.39425013e-01
1.88560297e-20 5.07582893e-01 -3.12479704e-02 7.30058131e-22
2.08259718e-03 1.08326852e-02 -6.22803436e-13]
===============================================================================
Optimization Results
===============================================================================
Optimal Coefficients
-------------------------------------------------------------------------------
c_d_fuselage_x: 0.002082597182204064
c_d_fuselage_y: 0.010832685190091253
c_d_fuselage_z: -6.228034359331524e-13
vertical_c_m_drag_z_lin: -0.03124797041333541
vertical_c_m_drag_z_quad: 0.5075828926443368
vertical_c_m_leaver_lin: 1.885602973253947e-20
vertical_c_m_leaver_quad: 0.33942501322805907
vertical_c_m_rolling: 7.300581311503944e-22
vertical_rot_drag_lin: 0.15906667162903768
vertical_rot_thrust_lin: -0.13827547174291024
vertical_rot_thrust_quad: 4.19292060688978

  -------------------------------------------------------------------------------
                              Optimization Metrics                               
  -------------------------------------------------------------------------------
  R2: 0.9876949185774294
  RMSE: 0.609447037495574
  
  -------------------------------------------------------------------------------
  Complete results saved to: 
  model_results/multirotor_model_2022-04-22-16-07-16.yaml

`

from data-driven-dynamics.

and figures




:

from data-driven-dynamics.

@AEGEAN99 Could you try with the "LinearRegressor" Instead of the "QPOptimiser"?

also might help if you updte to latest master

from data-driven-dynamics.

Actually I dont know how to change optimizer. Should I download the latest master PX4-Autopilot? Or should I install the version in your repo?

from data-driven-dynamics.

@AEGEAN99 Sorry I modified my initial response.

No, the latest master of this repo

from data-driven-dynamics.

okay I am trying now

from data-driven-dynamics.

@AEGEAN99 The plots remind me of a problem that I fixed a while ago: #193 Probably worth checking if the version you are using has the changes included

from data-driven-dynamics.

still not working,
Here is latest one with qpoptimizer

from data-driven-dynamics.

And here latest with linearRegressor

from data-driven-dynamics.

This happens much more slower

from data-driven-dynamics.

@AEGEAN99 Ah, now it is working.

Could you try

make estimate-model model=quadrotor_model log=resources/quadrotor_model_trajectory.csv

from data-driven-dynamics.

with linearRegressor?

from data-driven-dynamics.

@AEGEAN99 Yes, it is basically just swapping the log

from data-driven-dynamics.

It is working but has small oscillations like vibration

n

from data-driven-dynamics.

Here is the figures of the model. Are they satisfying?
And have you determined the problem? Thank you:DD

from data-driven-dynamics.

@AEGEAN99 Yes, so this means that it has successfully estimated the dynamics.

I am not sure why the vehicle is oscillating, but probably coming from the fact that the moment parameter estimation is not perfect.

from data-driven-dynamics.

But when I try the quadrotor_model.csv it agin crashes

from data-driven-dynamics.

What should I do for my own log files

from data-driven-dynamics.

@AEGEAN99 Yes, because it does not contain enough information - as you can see in the logs there is no angular excitation of the system and the system is unable to successfully estimate the correct parameters

Same for your own logs - You need a log that has enough information to estimate the parameters.

@manumerous Probably we should remove the degenerate logs?

from data-driven-dynamics.

Thanks for taking over Jay!

Yes I agree with you. I will quickly make a PR and put the trajectory log as our default log. That is indeed confusing.

Furthermore I think this issue points towards the fact that we need to improve the user feedback on the estimated parameters.

@AEGEAN99 as Jay already pointed out estimating the angular dynamics is usually harder than the linear ones. Can you try to fly some maneuvers where you have a lot of angular acceleration and try the pipeline again?

from data-driven-dynamics.

Adressed in #199

from data-driven-dynamics.

Thanks for reporting!
Feel free to close the issue if your problem is solved.

from data-driven-dynamics.