Overcooked V2 default observation view sizes and contents

[abaisero]

Hello,

I'm trying to understand how the view sizes in Overcooked V2 are calculated, and there's something that strikes me out as odd.
This is the shape function from the environment:

    def _get_obs_shape(self) -> Tuple[int]:
        if self.agent_view_size:
            view_size = self.agent_view_size * 2 + 1
            view_width = min(self.width, view_size)
            view_height = min(self.height, view_size)

It seems to take the smallest between (1) the agent's view size, and (2) the map's total size, and use that as the observation size.

I don't understand why that is? Why would the agent's view size be limited by the size of the layout? If an agent has view size of 2, it should have a view-size of 5. Even in a grid of size 4, there are situations (e.g., the agent being near the edge of the map) where the agent is unable to see the entire map. If the goal here was to take the smallest necessary viewsize, I think it should be min(2 * self.width, view_size) or something like that, but I'm not sure that's the intent here, and in general I think it's risky to overwrite a configured quantity like the agent view size anyway.

Is there any documentation on the design process of the overcooked v2 observations? both in terms of the sizes and the actual contents?

[tobiges]

The view size is calculated by view_size = self.agent_view_size * 2 + 1 the plus one is simply adding the cell of the agent itself which is added. If you configure an agent view size of 1, the agent will be able to observe its own cell and all surrounding ones. Then we take the smallest necessary view size by taking the min on the width/height and the view size. This is similar to what you pointed out with min(2 * self.width, view_size), however, the plus one is needed to account for the agents own cell. The agent view size could probably also be fixed to the agent view size regardless of the layout size; however, in most scenarios the agent view size will be much smaller than the map size.

[abaisero]

Hi tobiges, thanks for the reply;

I understand the view_size = self.agent_view_size * 2 + 1 part, that's not what I'm asking about, I'm only asking about the min part. I probably gave the wrong impression with my min(2 * self.width, view_size) comment (note that I wrote 2 * self.width, not 2 * self.agent_view_size) but I'm not sure I can easily explain it clearly, and it's probably irrelevant, so please ignore it.

You say "we take the smallest necessary view size" but I don't know what you mean by that, and you say " in most scenarios the agent view size will be much smaller than the map size", but why is that a problem? Could you elaborate?