Adventures in Machine Learning for Robotics

[LHACK4142]

One of my teammates in my FTC robotics team suggested using AI to drive the robot instead of painstakingly pre-programming instructions. I thought this was kind of far-fetched, but as I researched the idea, it seemed more and more plausible, so we're going to give it a go! If we can get it working, it could honestly be a bit of a game-changer; basically nobody else is doing it, and it could save so much time and accelerate development dramatically.

I want to document the process so that we can write about it in our essays and stuff, and also so that we can maybe make a tutorial, so I'll be posting whatever stuff I found here.

So far, I've decided that I want to use a reinforcement learning model because they use agents in environments, which is easily adapted to a robot in a playing field. I'm not sure Q-learning would work because from what I gather, it fills out a table of all possible states and decides the best thing to do based on that table, and we'd have around 2^140 states. Therefore, I think I'll use deep learning. Since I need to use Java for FTC, I think it's time to start learning deeplearning4j!

[LHACK4142]

Deeplearning4J's reinforcement learning model sucks. The FRC competition (which is cooler bigger FTC) uses python, which I know, so I think I'll be using Python actually.