Update metrics.py

[TachikakaMin]

Bug

obstacles is local information

traversable_cells = np.size(obs['obstacles']) - np.count_nonzero(obs['obstacles'])

[TachikakaMin]

@Tviskaron

Hi Alexey,

Please check this bug

[Tviskaron]

Hi Yimin,

Could you explain why this is a bug? The wrapper calculates density based on the agent’s field of view. The number of global obstacles and the number of free cells are constants.

[TachikakaMin]

Hi Yimin,

Could you explain why this is a bug? The wrapper calculates density based on the agent’s field of view. The number of global obstacles and the number of free cells are constants.

It seems if there is no obstacle, obs['obstacles'] will be [].

traversable_cells will be 0

[Tviskaron]

obs['obstacles'] is a boolean matrix, so its size should not be empty. However, to use the wrapper, you need to set observation_type to either POMAPF or MAPF.

Here is a small example where I tried to reproduce the wrapper behaviour:

from pogema import pogema_v0, GridConfig
from pogema.wrappers.metrics import AgentsDensityWrapper
import numpy as np

grid = """
......
......
..a...
...A..
......
"""
env = AgentsDensityWrapper(pogema_v0(grid_config=GridConfig(map=grid, observation_type='POMAPF', obs_radius=2)))
obs, info = env.reset()

env.render()

print('Obstacles matrix size:', np.size(obs[0]['obstacles']))

[TachikakaMin]

It seems if I change add_artificial_boarder to be false, the obs['obstacles'] will become empty. When I run pogema_benchmark.

[Tviskaron]

global or local ones? Could you provide the minimal code to reproduce the problem?