Option for full observability about minigrid HOT 12 CLOSED

I guess I should point out, if you want a fully observable setting, the quickest "hack" would be to just use the RGB image from the human render, you can produce this with img = env.render('rgb_array').

from minigrid.

A lazy way of doing it would be something like this?

class FullyObsWrapper(gym.core.ObservationWrapper):
    """
    Fully observable gridworld
    """

    def __init__(self, env):
        super().__init__(env)
        self.__dict__.update(vars(env))
        self.observation_space = spaces.Box(
            low=0,
            high=255,
            shape=(self.env.grid_size * 32, self.env.grid_size * 32, 3),  # number of cells
            dtype="uint8"
        )

    def observation(self, obs):
        if self.env.grid_render is None:
            return np.zeros(shape=self.observation_space.shape)  # dark screen as init state?
        return self.env.grid_render.getArray()

But how do you encode the direction of the agent? Maybe a tiled vector over the last dimension, so you will get the total obs as 800x800x4?

from minigrid.

If you render the whole grid that way and get the pixel data, then the agent will be visually rendered in there already.

from minigrid.

You'd have to encode the agent in the observation differently, but it seems very doable. Are you volunteering? ;)

from minigrid.

By full observability you mean the whole grid-world, the objects, position and orientation of the agent right?

from minigrid.

Yes, a full view of the grid including the agent.

from minigrid.

Had the same question about rendering the whole grid. Using img = env.render('rgb_array') was a nice hack. Thanks for the suggestion.

from minigrid.

@maximecb there is a little issue about attributes sharing. The env has more attribute than the gym env, thus wrapping requires the attributes update at least for testing purposes, i don't think you use them in the standard setting.

Shall I keep it like this and forget for now, or try to refactor a little bit and submit a PR?

from minigrid.

@d3sm0 can you clarify what the issue is? May be nice to get a PR :)

from minigrid.

@maximecb sure. The issue is that when you wrap some env in a gym.Wrapper instance, you loose access to the attributes of the env. Thus it doesn't pass the test unless you don't share the missing attributes with self.__dict__.update(vars(env)) and your tests breaks.

I'm working on #13 and PR :)

from minigrid.

Sweet. Thanks @d3sm0 !

from minigrid.

Working solution contributed by @d3sm0 in #24

from minigrid.