Self-Driving Car Engineer Nanodegree Program

• cmake >= 3.5
• All OSes: click here for installation instructions
• make >= 4.1
  • Linux: make is installed by default on most Linux distros
  • Mac: install Xcode command line tools to get make
  • Windows: Click here for installation instructions
• gcc/g++ >= 5.4
  • Linux: gcc / g++ is installed by default on most Linux distros
  • Mac: same deal as make - [install Xcode command line tools]((https://developer.apple.com/xcode/features/)
  • Windows: recommend using MinGW
• uWebSockets
  • Run either ./install-mac.sh or ./install-ubuntu.sh.
  • If you install from source, checkout to commit e94b6e1, i.e.

    git clone https://github.com/uWebSockets/uWebSockets 
    cd uWebSockets
    git checkout e94b6e1
    

    Some function signatures have changed in v0.14.x. See this PR for more details.
• Simulator. You can download these from the project intro page in the classroom.

There's an experimental patch for windows in this PR

1. Clone this repo.
2. Make a build directory: mkdir build && cd build
3. Compile: cmake .. && make
4. Run it: ./pid.

The P.I.D parameters are first chosen by hand. I first try "Kp=0.4; Ki=0; Kd=2" and found that it had many wiggling effects. This is because the large Kp coefficient makes overshooting. Then I reduced the Kp value and set Kp=0.2 as the manual result.

In addition, I found that it is hard to tune Ki because the integral error is typically larger than proportional error in hundreds of magnitude. So I just set it to zero and left it be tuned by twiddling function.

In order to evaluate each "Kp,Ki,Kd" setting, we need to restart the simulator. By sending "42[\"reset\",{}]" to WebSocket, simulator is able to restart. The twiddling function can then test the performance of each setting and found the best results.

Each perforamnce evaluation is tested in time stamps 3000 (#define TWIDDLE_TIME_NUM 3000 in program), which is about 59 seconds. Then twiddling function will update the parameters and restart simulator.

The initial setting is found by manually tunning: "Kp=0.4; Ki=0; Kd=2", and I set the initial step size as:

twiddle_step_size[0] = 0.01; // pd[0] in course lecture
twiddle_step_size[1] = 0.0001; // pd[1] in course lecture
twiddle_step_size[2] = 0.01; // pd[2] in course lecture

Moreover, the tolerance, tol, is set to 0.01, and if "cur_tol<tol", the twiddling function converges. One thing that is different from the course is the calculation of cur_tol. Instead of using sum(pd) as in the course lecture, I used:

double cur_tol = twiddle_step_size[0]+twiddle_step_size[1]*100+twiddle_step_size[2];

I scaled 100 times of twiddle_step_size[1] (pd[1]). This is because the integral error is typically larger than proportional error in hundreds of magnitude.

As we can see in the following figure, cur_tol converges toward tol in the 84th iteration. (about 84 mins)

The 84th iteration got the following results:

double Init_Kp = 0.211;
double Init_Ki = 0.0003189;
double Init_Kd = 2.00109;

It can successfully drive a lap around the track!