AI with a EV3 Legomindstorm

In my graduation thesis I wrote about Artificial Intelligence and started to explore a little bit what machine-learning is all about. For the practical part I have taken up the challenge to imitate the project "BrickClassifi3r " from CT Magazine. The goal is to make a Lego-Mindstorm-EV3 robot learn to distinguish between three different objects using an infrared sensor and a neural network. The neural network was just a very simple 3-layer network (24-6-3) with one input-, one output- and one hidden-layer. The training happens with a simple Python script and the TensorFlow environment. Therefore, I collected 250 training sets and 125 test sets per object. Unfortunately, after running the script (doing the training) I could only achieve a maximum hit rate of 59%, but this could also have been due to the quality of my data sets. In general, it was an amazing opportunity to explore what AI is on a very basic level.