def _s_to_obs(self, s):
"""Returns an array representation of the env to be used as an observation.
The representation has dimensions (height, width, 6) and consist of:
- 2d grid with 1 in the agent's position if it does not carry a battery and
2 if it does, and 0 everywhere else
- 2d grid with life of each train in the train's position
- one-hot encoding of all trains with 0 life
- one-hot encoding of all batteries that have not been collected
- one-hot encoding of all batteries that have been collected
- one-hot encoding of the goals
"""
layers = [
[s.agent_pos],
[s.train_pos] if s.train_life > 0 else [],
[s.train_pos] if s.train_life == 0 else [],
[pos for pos, present in s.battery_present.items() if present],
[pos for pos, present in s.battery_present.items() if not present],
self.feature_locations,
]
obs = get_grid_representation(self.width, self.height, layers)
if s.carrying_battery:
obs[:, :, 0] *= 2
obs[:, :, 1] *= s.train_life
return np.array(obs, dtype=np.float32)
def _s_to_obs(self, s):
"""Returns an array representation of the env to be used as an observation.
The representation has dimensions (height, width, 6) and consist of:
- 2d grid with the agents orientation at the agent's position and 0 else
- 2d grid with 1 in the agent's position iff it is carrying an apple, and
0 everywhere else
- one-hot encoding of all trees that have apples
- one-hot encoding of all trees that do not have apples
- one-hot encoding of all buckets
- 2d grid with the number of apples for each bucket in the bucket's location
and 0 everywhere else
"""
orientation, agent_x, agent_y = s.agent_pos
layers = [
[(agent_x, agent_y)],
[(agent_x, agent_y)] if s.carrying_apple else [],
[pos for pos, has_apple in s.tree_states.items() if has_apple],
[pos for pos, has_apple in s.tree_states.items() if not has_apple],
self.bucket_locations,
self.bucket_locations,
]
obs = get_grid_representation(self.width, self.height, layers)
obs[:, :, 0] *= orientation + 1
for (bucket_x, bucket_y), num_apples in s.bucket_states.items():
obs[bucket_y, bucket_x, 5] = num_apples
return np.array(obs, dtype=np.float32)
def _s_to_obs(self, s):
"""Returns an array representation of the env to be used as an observation.
The representation has dimensions (5, height, width) and consist of one-hot
encodings of:
- the agent's position
- intact vases
- broken vases
- carpets
- goals
"""
layers = [
[s.agent_pos],
[pos for pos, intact in s.vase_states.items() if intact],
[pos for pos, intact in s.vase_states.items() if not intact],
self.carpet_locations,
self.feature_locations,
]
obs = get_grid_representation(self.width, self.height, layers)
return np.array(obs, dtype=np.float32)
def _s_to_obs(self, s):
layers = [[s]]
obs = get_grid_representation(self.width, self.height, layers)
return np.array(obs, dtype=np.float32)