Juan Simó Álvarez

Robotics Software Engineer

2 minute read

Progress This Week

Objective

The objective, as stated, was to find a way to get rid of the currupted data and train a nueral and specalist pilots

Currupted data

After trying to find the way to test the datasets and find the wrong data I realized it would be faster just to re-make the datasets given the results I was obtaining. However I first tried to experiment on diferent combinations of the actual datasets and ended up with a pilot that worked quite well if I avoided using the dizzy datasets and worked a bit the afinning. With that and some new data I advanced to the specialists.

Specialists

In order to compare the Double Neural pilot with the simple one I needed to use the same datasets and to make sure the sum of the data that went into the specialists added up to the amount of data in the neural pilot. After dividing the datasets using the parameters that best worked last time and using the same datasets for training the specialists that the ones used in the Neural Pilot training I got both the experts and tested them. They did good in most circuits, but there where 2 main problems:

  1. They were not doing Nurburing, our main circuit for testing given that is a complex one and that there is no dataset of it in the trainings, consistently.
  2. The classifier was calssifying too many curves when used on the pilot to decide wich net to ask.

To change this I developped a more aadvance clasifier using open cv. As a result I ended up discarding the Montreal circuit given taht low definition of the line was causing problems on the clasification of the images. Of course I tried first to adapt the classifier to make it able to classify the Montreal data correctly, but after the comparison of the nets this proved to be less reliant than the alternative.

However, after having remooved one of the circuits from the datasets I had less data and to re-train the Neural Pilot for the reasons already stated, if I want to compare them they have to be trained on equal terms.

This time, not wanting to spare resources, I tripled the data given and got great results:

  • The specialists did consistently good in every circuit.
  • The neural net did good in every circuit, but prooved to have worse capabilities on recovering, specially in straight lines where there is less data on recoverage.

Next week

For the next week I plan on having the metrics and thus to make a final post with the result of the comparisons, videos of both nets and some pictures on how everything works.