Does changing the observation space seems to require changing C code?

[dmadeka]

Right now, the observation_keys argument is compared against OBSERVATION_DESC (https://github.com/facebookresearch/nle/blob/master/nle/nethack/nethack.py#L145) which is pinned against the set_buffers in (https://github.com/facebookresearch/nle/blob/master/win/rl/pynethack.cc#L126)

The gym observation space is also defined from observation_keys. So whats the intended design for adding something to the observation space like an occupancy map?

The current solution looks like (is this the intended design?):

class NetHackOccupancyMap(NetHackScore):
    def __init__(self, *args, **kwargs):
            super().__init__(*args, **kwargs)
            self.positions_visited = {(0, 1):
                        np.zeros((21, 79))
                    }

            occupancy_map_tuple = ("occupancy_map",
                                gym.spaces.Box(low=0, high=nethack.MAX_GLYPH, **nethack.OBSERVATION_DESC["glyphs"]))

            space_dict = kwargs.pop("space_dict", None)
            if not space_dict:
                space_dict = dict(NLE_SPACE_ITEMS + (occupancy_map_tuple,))
            else:
                space_dict = dict(space_dict + (occupancy_map_tuple,))
    
            observation_keys_new = kwargs.pop("observation_keys", (
                "glyphs",
                "chars",
                "colors",
                "specials",
                "blstats",
                "message",
                "inv_glyphs",
                "inv_strs",
                "inv_letters",
                "inv_oclasses",
                "screen_descriptions",
                "tty_chars",
                "tty_colors",
                "tty_cursor",
                "occupancy_map",
            ))
        
            self.observation_space = gym.spaces.Dict(
                {key: space_dict[key] for key in observation_keys_new}
            )
          

[cdmatters]

Hi @dmadeka ,

Thanks for opening the issue. You seem to want to extend NLE with a new observation and a new task. I will ignore the latter as there are couple of examples you can check in nle/env/tasks.py, and just address the observation.

If you wish to add a new observation you need to first ask what are you trying to observe. Is this a new piece of state that carries new internal information out of NetHack? Or is this a new observation derived from existing state that has already been exposed by the NLE.

If the former, I suggest looking at the recent commit that added the misc observation to NLE. Here is the commit and PR.

If the latter, then I suggest writing a wrapper (or simply a new task) that wraps around that intercepts the step and reset functions to add more observations. I'm sure there are examples of this elsewhere online since you would just need to follow the gym API.

If this answers doesn't answer your question feel free to reopen the issue.

---

( Finally on a personal note @dmadeka - I want to thank you for you contributions, you are very close to merging your contribution to NLE (PR #202 )! All you need to do is pass the linting tests! )

[dmadeka]

Reopened #217 correctly and fixed the linting issues