[elpidiovaldez5]

I like experimenting with Reinforcement Learning.   Although there are excellent highly sophisticated environments available such as OpenAi gym, I find that I often want to set up a very simple 'toy' grid world problem that presents a very specific learning challenge. 

Does anybody know of an 'Environment Builder' that can be used to rapidly define 2D worlds with simple properties ?

Ideally an Environment Builder would allow:

• Random (according to some statistical parameters) generation and placement of items in the environment.
• Labyrinth generation
• Support movement actions and collision detection
• Support 1st and 3rd person agent viewpoint
• Allows actions and responses to be defined for items in the world
• Can provide a GUI view of the world as the agent moves in the environment
• Easy integration with RL/Deep Learning environment (tensorflow in my case)

   
I imagine the Environment Builder being a framework, or just a library, in some interactive language (such as python).

[LOCKSUIT]

Blender. The Game Engine inside Blender. Also check out UPBGE - a team is/was worked on a even more upgraded version of Blender's game engine.

I've used Blender quite a lot. It's very easy to use too or maybe I just spent too long in it haha. It's great for rapid testing and all you say, wonderfully easy. However, while BGE/UPBGE is great and perfect in some areas, it isn't top notch as it could be, sorta like a bummy program, but not necessarily bummy either haha. UE4 is a power horse, it can do like, anything, unless I'm wrong. And excellent physics etc omg. Just, UE4 has a learning curve pft! No inuitiveity there.