Sunday, April 28, 2013

Final portfolio


Portfolio Final Chapter


Lectures
  • I found the lectures very difficult to follow.The topics grew very complex very quickly, and I didn’t feel like there were enough examples to follow how and when to implement the formulas.

Polls
  • The polls would have been a great idea if they were updated more consistently. I ended up having to check at random times when they were not up before class, and if I was away from my computer for a couple days as I was sometimes on weekends, and the poll was put up late, I inevitably missed some. I also tended to just guess if I didn’t know the answer opposed to trying to look it up, because they were not graded for accuracy. To help with this I would recommend in the future giving maybe 0.4% for attempting the poll and the other 0.1% being for correctness, so there isn’t a huge punishment for getting it incorrect, but there is an incentive to look it up if you can.

Assignments
  • What I learned from the assignments, and from my peers:
    • The group projects were where most of the real learning happened in the course. When each of us knew and understood different aspects of the topic we were trying to implement, we were able to teach each other far more effectively than any lecture. The non-group projects felt like they were teaching a specific instance of an algorithm, and thats all I learned from them.
  • Favorite assignment:
    • My favorite assignment would have been the Quadcopter assignment, because of how interesting it was, but the time restraints giving us only 2-3 days for each part made it more stressful and grueling than enjoyable. As it stands my favorite assignment was the Q-learning one because we had more than adequate time to do it, it was not that difficult, and I learned something new and possibly useful for later projects.
  • Self assessment on assignments:
    • I did well on the assignments, but had trouble with a lot of them causing me to spend a great deal of time working on them. The one I had the most trouble with was the manipulator assignment because it was a really rough week for other classes, and even after implementing the algorithm, the proportions of the conversion back and forth from pixels to units had to be tweaked dozens of times before it worked.



Quizzes
  • What I thought of the quizzes:
    • I felt the quizzes were too common and not worth enough of the grade to be able to make time to consistently study for, but were too difficult to reliably do well on without spending a good deal of time preparing.  It felt like they just added an additional level of stress to the course. Multiple choice was the best format for the quizzes, because it allowed recognition by seeing the answer, which helps when there is little to no prestudying.
  • Self assessment on quizzes:
    • I did well on the earlier quizzes, but did not do as well on later quizzes as the material got more complicated. Studying for a quiz worth 3% of my grade just never took priority of working on assignments and studying for large exams in other classes.

Portfolio
  • I believe the portfolio did little to nothing to actually help me or anyone else I talked to understand the topics. The only thing it felt like it did was add something else to the to do list for the week after completing the corresponding assignment.

Final Comments
  • I felt like in the course we covered too much too quickly. I would have liked to go back and study some the of more complex topics like computer vision of path finding in more detail opposed to keeping moving forward at so brisk a pace that it felt like I was getting an overview of everything without learning anything. The topics we covered affirmed my interest in in the particular topics of software AI and computer vision, while also teaching me that I would like to avoid the headache of working with robots again in the future.

Thursday, April 25, 2013

Assignment 6


Assignment 6 was interesting. My group chose option two and found two random slam algorithms from the same author, and an example data set. But due to the fact that so many of the variables were created by random third parties opposed to the instructors or students; we had no idea what to expect as output. My group ended up with very minimal data and found it very difficult to compare. For example one of the outputs was a nondescript array of ints, that after much testing we were able to determine were (x,y) coordinates listed in a 62 by 1 array that we had to pull into excel and divide up to graph. The assignment did however teach us how important SLAM algorithms are for analyzing data of the input of a run.

Assignment 5

Assignment 5 was interesting but had some flaws:

  • We had way more time than we needed for part one, but we were strapped for time on part two, and found it almost impossible to fit in the time to do part three in the last couple days before spring break.
  • The robot controls were not accurate, the quad-copter did not float evenly when told to just levitate, and had a hard time orienting itself to the visual target.
  • camera didn't always find the colors. We had to have the Quad-copter at a specific angle and distance to see all 3 circles.
  • too many graphs needed for part two these took several hours to make and many of them were not relevant.
Overall though Assignment 5 was not that bad. The quad-copter was much easier to work with than the Rovio.

Tuesday, March 12, 2013

Assignment 4

The "group" part wasn't hard we had some problems understanding what was being asked, but nothing serious.
The first part of the individual portion involved the robot so could not actually be done alone without trading the robot around. So we all had to meet up then wait on each other to take their turns with the robot. Then it was very difficult to get the data for the graphs. then the actual creation of the graphs took several hours. I didn't really learn anything but got very frustrated with the project.

Monday, February 25, 2013

Assignment 3

Assignment 3 I learned:
  • always assume Murphy's law do not wait until last minute.
  • do math in pixels if you have to draw it eventually.
  • there is a lot of weird displacement with how processing sees the mouse.
  • simulations are frustrating.

Friday, January 25, 2013

3.1 Blog write-up for Assignment 1

I'n the first assignment I learned quite a bit about robotic movement and perception:

  • When following a line always stop to turn, or you will end up far enough dis-aligned that PID cannot auto correct.
  • Make sure to put a delay on state changes, else the sensors will cause the states to change rapidly before actually completing the step, causing the robot to fail.
  • Sensors are not infallible. put in a margin of error.
  • do not try to implement the first solution you come up with before exploring less complicated possibilities; unless you want to spend hours creating a self correcting data structure, and adding a library for it.
  • simulation and real life are not the same, and the robot WILL act differently than it did in simulation with the same code. Make sure to test.
That is what comes to mind when I think about what I learned from the first assignment. Everything else is in the other write-up.

Tuesday, January 22, 2013

2.1
Write-Up Component
Create a diagram of the FSM in SimpleFSM. Go through the steps the FSM performs and
draw each state as a node. Connect nodes to each other with triggers.





4.1
Write-Up Component
For this portion of the write up, just include a screen shot of a successful compile of
Polu - louSimpleFSM.ino
5.1
Write-Up Component
Go through the robot specific code and look for the function which defines the PID controller
for the robot. Write up a quick description of what each of the components (error, integral,
derivative) contribute to the final control input. How would adjusting these values change
the behavior of the robot?

the guts of the PID function is as follows:

error/100 + integral_/50000 + derivative*3/10;
So, this gives us 3 variables that we are adjusting in order to determine their effects on the behavior of the robot, error, integral, and derivative. Proportion should be bulk of term, under-proportion results in control actions that are too small to be effective, large numbers result in an unstable system. ( Think over adjusting self-destruction ). Integral controls the speed at which the device resets, so a large value here can mean that the device continually overshoots when it tries to recenter Derivative is used to clean up the overshooting performed by the integral.

7.1
Write-Up Component
Give a detailed description of your learning algorithm. Why did you decide on the method
you implemented? How does it work in simulation? We have provided two dierent mazes
to test your learner on. Please provide screen shots of the resulting outputted FSM string
created by your learner for each maze.

We implemented a sensor driven depth first search generating a data structure that will contain the correct path. By senor driven I mean we don’t turn on delay we turn till we see a line in front of us. the data structure is a node tree with 4 possible pointers, 1 in each direction. It does work in simulation on both mazes.