def skill_stopping_func(self, path):
# Stop if I'm on target OR if I don't move (less useless moves = shorter training)
moves = np.array([[-1, 0], [0, 1], [1, 0], [0, -1]])
last_pos = path['observations'][-1][:2]
if len(path['observations']) > 1 \
and np.array_equal(last_pos, path['observations'][-2][:2]):
return True # I don't move -> stop skill
a = special.from_onehot(path["actions"][-1])
last_move = moves[a]
return np.array_equal(last_pos + last_move, self.target)
def unflatten(space, obs):
if isinstance(space, gym.spaces.Box):
return np.asarray(obs).reshape(space.shape)
elif isinstance(space, gym.spaces.Discrete):
return special.from_onehot(obs)
elif isinstance(space, gym.spaces.Tuple):
dims = [flat_dim(c) for c in space.spaces]
flat_xs = np.split(obs, np.cumsum(dims)[:-1])
return tuple(unflatten(c, xi) for c, xi in zip(space.spaces, flat_xs))
else:
raise NotImplementedError
def unflatten(self, x):
return special.from_onehot(x)