The NumGrid environment consists of a grid of hand-written digits images loaded from a MNIST database file in IDX gzipped format, as can be found on LeCun's website.
The environment holds a cursor representing the agent's local view on the world (aka the grid); the cursor can either be moved on a small distance in one of the 4 orthogonal directions, or be directly teleported at a given position (the latter constituting a substantially larger action space).
The agent's goal is to reach the highest possible speed in accurately guessing the digit it is currently viewing. Right labelling leads to a positive reward, and wrong labelling to a negative one. The agent can of course take some steps to prepare its answer by exploring the image, in which case it can label with a 10 to tell the environment to ignore the answer.
cd gym-numgrid
pip install -e .To get started with the environment, you can run it with one of the agents in examples/agents. An example test loop is provided in examples/test.py:
python -m examples.testThis will work out of the box only if you downloaded the MNIST training data (files train-images-idx3-ubyte.gz and train-labels-idx1-ubyte.gz) in the directory you launched the script from. You can specify different paths in the environment's constructor parameters.