louism33 / gym-minigrid

Minimalistic gridworld package for OpenAI Gym

With thanks to Maxime Chevalier-Boisvert (maximecb), author of the original https://github.com/maximecb/gym-minigrid

This fork of https://github.com/maximecb/gym-minigrid adds an environment for the AI Safety concern of Scalable Supervision in this file. It also adds a wrapper for partial observability, and some basic rendering in rendering and minigrid.

This forms the environment that I benchmarked agents against in my thesis.

for a quick start on the SS environment, run ./manual_control.py --env MiniGrid-ScalableOversightVISUALISE-v0

More details are the file and the thesis.

You will need to use this in conjunction with https://github.com/louism33/torch-ac and https://github.com/louism33/gym-minigrid

This is an example of an agent performing on the training dataset generated by the hyperparameters Phones=1, Dirts=1, Known=1, Unknown=1. It presses the button when the situation is Unknown (meaning it does not know what is under the grey Mess/Unknown cells), and does not press it on the Known situations, when the Mess cells always have the same cells underneath.