Skip to content

Latest commit

 

History

History
79 lines (55 loc) · 4.21 KB

WRITEUP.md

File metadata and controls

79 lines (55 loc) · 4.21 KB
Raw

Flying Car Engineer Nanodegree - Project 4: Estimation

In this write-up I will describe how I addressed each of the steps outlined in the project README.

Throughout this document I will be referencing the provided Estimation for Quadrotors as EfQ.

1: Sensor Noise

I run scenario #6 in the simulator, which generates some sensor data from a static quad, in particular the GPS and accelerometer x measurements. The sensor data were saved in 2 txt files which I then processed (using a Jupyter notebooks and Pandas) and figured out the standard deviation of the two measured signals as follows:

  • std(Quad.GPS.X) = 0.709497
  • std(Quad.IMU.AX) = 0.496512

Finally, I plugged these values in config/6_Sensornoise.txt and run the simulator and run the simulator again. This time, around 68% of the measurements fall within the +/- 1 standard deviation range, as expected for a Gaussian noise model).

Simulator Scenario #6

2: Attitude Estimation

I modified the function UpdateFromIMU() in QuadEstimatorEKF.cpp. In particular, I implemented a better integration rate gyro attitude scheme by following EfQ section 7.1.2 and the instructions in the provided code hints about the Quaternion class.

As a result, the attitude errors were limited within 0.1 rad for each Euler angle.

Simulator Scenario #7

3: Prediction Step

Part 1

I implemented the state prediction step in the PredictState() functon in QuadEstimatorEKF.cpp, as per EfQ equation (49). Now in simulator scenario 08_PredictState the estimated state reasonably tracks the actual state.

Simulator Scenario #8

Part 2

I calculated the partial derivative of the body-to-global rotation matrix in the function GetRbgPrime() in QuadEstimatorEKF.cpp, as per EfQ equation (52). Then, I implemented the covariance predictin step in Predict(), as per line 4 of EfQ Algorithm (2), also using EfQ equation (51). Finally, I tuned QPosXYStd and QVelXYStd parameters in QuadEstimatorEKF.txt so as to capture the magnitude of the error in simulation scenario 09_PredictionCov.

Eventually, in in simulation scenario 09_PredictionCov our covariance (the white line) grows reasonably like the data and the charts resemble the provided examples of a 'good' solution.

Simulator Scenario #9

4: Magnetometer Update

I implement the function UpdateFromMag() as per EfQ equations (56) and (58) and tuned the parameter QYawStd. As a result, my estimator met the success criteria for smulation scenario 10_MagUpdate.

Simulator Scenario #10

5: Closed Loop + GPS Update

I switched to using my estimator and a realistic IMU by setting Quad.UseIdealEstimator = 0 in 11_GPSUpdate.txt and commenting out the rrelevant lines in 11_GPSUpdate.txt. Then I implemented the function UpdateFromGPS() as per EqF equations (53) and (55) and tuned the GPS measurement std deviations in QuadEstimatorEKF.txt. As a result, my estimator met the success criteria for simulation scenario 11_GPSUpdate

Simulator Scenario #11 - Provided controller

6: Adding Your Controller

I replaced the provided controller and associated control parameters with the controller I developed for project 3 and its own control parameters. I tried running scenario 11 again but the drone crashed. Then I re-tuned my controller's parameters, starting with the body rates. Finally, my controller - estimator combo was able to meet the success criteria for scenario 11, even though the trajectory looks a bit ugly :)

Simulator Scenario #11 - My controller from Project 3

Back to back runs that PASS