Minigrid VS ma-gym

If you want something as simple as possible, I'd go with MiniGrid, and if you want to have a richer world with more complex settings, then MiniHack.

This is a pretty hard domain to start with as someone "brand new" to AI. If you're interested in the vision aspect, I'd suggest you start by training a DNN for the CIFAR-10 task. There are plenty of tutorials out there. If you're more interested in the navigation aspect, you could start by training a Q-learning agent to solve some of the simpler problems in gym-minigrid.

Hello guys I tried to build a custom environment using maxicymeb repo

For learning and experimentation with RL algorithms, I suggest using a grid world implementation: observations are simple enough (most implementations have a one-hot layered observation) that you do not need deep conv layers to learn complex visual features. You can also make grid worlds as simple or as complex as you like by adding enemies, objects, key-door pairs, changing the size of the grid or decreasing observation radius, etc. There is a reason they are commonly used in research.

Take a look at: https://github.com/maximecb/gym-minigrid

Of course I conducted research (or rather... searched on google :)) before making this post. I found a simple third party gym environment that implements multi-agent rl called 'ma-gym' (https://github.com/koulanurag/ma-gym) but I'm not quite sure how I would train two Agents and let them play inside one environment.

There are projects working on a multi agent version of gym. Maybe you find some inspiration there.