def find_good_goal(self):
b = self.items_bytype['pushable_block'][0]
mincost = dut.get_big_cost()
if self.direction == 'up':
for i in range(self.mapsize[0]):
p, cost = dut.dijkstra_touch_cost(self, b.attr['loc'],
(i, self.mapsize[1] - 1))
if cost < mincost:
mincost = cost
goal_loc = (i, self.mapsize[1] - 1)
minp = p
elif self.direction == 'down':
for i in range(self.mapsize[0]):
p, cost = dut.dijkstra_touch_cost(self, b.attr['loc'], (i, 0))
if cost < mincost:
mincost = cost
goal_loc = (i, 0)
minp = p
elif self.direction == 'right':
for i in range(self.mapsize[1]):
p, cost = dut.dijkstra_touch_cost(self, b.attr['loc'],
(self.mapsize[0] - 1, i))
if cost < mincost:
mincost = cost
goal_loc = (self.mapsize[0] - 1, i)
minp = p
else:
for i in range(self.mapsize[1]):
p, cost = dut.dijkstra_touch_cost(self, b.attr['loc'], (0, i))
if cost < mincost:
mincost = cost
goal_loc = (0, i)
minp = p
return mincost, goal_loc, minp
def get_supervision(self, featurizer):
states = []
actions = []
# this is greedy, not optimal:
goal_remaining = True
while goal_remaining:
mincost = dut.get_big_cost() + 1
goal_remaining = False
for j in range(self.ngoals_active):
aloc = self.agent.attr['loc']
g = self.items_bytype['goal'][self.goal_order[j]]
gloc = g.attr['loc']
v = g.attr.get('_visited')
if v is None or v is False:#why is v False?
goal_remaining = True
p, cost = dut.dijkstra_touch_cost(self, aloc, gloc)
if cost < mincost:
mincost = cost
mingloc = gloc
minp = p
if mincost >= dut.get_big_cost(): # necessary goal is unreachable
return [[featurizer.featurize(self), 'stop']]
else:
path = dut.collect_path(minp, mingloc)
thisgoal_actions = dut.path_to_actions(path)
actions.extend(thisgoal_actions)
for a in thisgoal_actions:
states.append(featurizer.featurize(self))
self.act(a)
self.update()
return list(zip(states, actions))
def go_and_push(self, next_block_loc, actions, states):
b = self.items_bytype['pushable_block'][0]
pl = self.get_push_location(b.attr['loc'], next_block_loc)
p, cost = dut.dijkstra_touch_cost(self, self.agent.attr['loc'], pl)
if cost < dut.get_big_cost():
path_to_push = dut.collect_path(p, pl)
else:
return actions, states, False
to_push_loc = dut.path_to_actions(path_to_push)
actions.extend(to_push_loc)
for a in to_push_loc:
states.append(self.featurizer.featurize(self))
self.act(a)
self.update()
if pl[0] < b.attr['loc'][0]:
a = 'push_right'
elif pl[1] < b.attr['loc'][1]:
a = 'push_up'
elif pl[0] > b.attr['loc'][0]:
a = 'push_left'
else:
a = 'push_down'
actions.append(a)
states.append(self.featurizer.featurize(self))
self.act(a)
self.update()
return actions, states, True
def get_supervision(self, featurizer):
a = self.agent
b = self.items_bytype['pushable_block'][0]
p, cost = dut.dijkstra_touch_cost(self, b.attr['loc'], self.goal_loc)
states = []
actions = []
if b.attr['loc'] == self.goal_loc:
return [[featurizer.featurize(self), 'stop']]
if cost < dut.get_big_cost(): # switch is not blocked
path = dut.collect_path(p, self.goal_loc)
path = path[1:]
for loc in path:
pl = b.get_push_location(loc)
p, cost = dut.dijkstra_touch_cost(self, self.agent.attr['loc'],
pl)
if cost < dut.get_big_cost():
path_to_push = dut.collect_path(p, pl)
else:
return [[featurizer.featurize(self), 'stop']]
to_push_loc = dut.path_to_actions(path_to_push)
actions.extend(to_push_loc)
for a in to_push_loc:
states.append(featurizer.featurize(self))
self.act(a)
self.update()
if pl[0] < b.attr['loc'][0]:
a = 'push_right'
elif pl[1] < b.attr['loc'][1]:
a = 'push_up'
elif pl[0] > b.attr['loc'][0]:
a = 'push_left'
else:
a = 'push_down'
actions.append(a)
states.append(featurizer.featurize(self))
self.act(a)
self.update()
return list(zip(states, actions))
def get_supervision(self, featurizer):
gloc = self.goal_loc
p, cost = dut.dijkstra_touch_cost(self, self.agent.attr['loc'], gloc)
if cost >= dut.get_big_cost():
return [[featurizer.featurize(self), 'stop']]
path = dut.collect_path(p, gloc)
actions = dut.path_to_actions(path)
states = []
for a in actions:
states.append(featurizer.featurize(self))
self.act(a)
self.update()
return list(zip(states, actions))
def get_supervision(self, featurizer):
states = []
actions = []
for i in range(self.ngoals_active):
current_goal = 'goal' + str(self.goal_order[i])
for j in self.items_bytype['goal']:
if j.attr['@goal'] == current_goal:
gloc = j.attr['loc']
p, cost = dut.dijkstra_touch_cost(self, self.agent.attr['loc'], gloc)
if cost >= dut.get_big_cost():
return [[featurizer.featurize(self), 'stop']]
else:
path = dut.collect_path(p, gloc)
thisgoal_actions = dut.path_to_actions(path)
actions.extend(thisgoal_actions)
for a in thisgoal_actions:
states.append(featurizer.featurize(self))
self.act(a)
self.update()
return list(zip(states, actions))
def get_supervision(self, featurizer):
gloc = self.items_bytype['goal'][0].attr['loc']
aloc = self.agent.attr['loc']
p, cost = dut.dijkstra_touch_cost(self, aloc, gloc)
states = []
actions = []
if cost < dut.get_big_cost(): # goal is on this side...
path = dut.collect_path(p, gloc)
actions = dut.path_to_actions(path)
for a in actions:
states.append(featurizer.featurize(self))
self.act(a)
self.update()
else: # gonna have to move the block.
#find where to push:
bloc = self.items_bytype['pushable_block'][0].attr['loc']
dw = round((aloc[0] - bloc[0]) / (abs(aloc[0] - bloc[0]) + .01))
dh = round((aloc[1] - bloc[1]) / (abs(aloc[1] - bloc[1]) + .01))
bw = bloc[0]
bh = bloc[1]
if self.is_loc_reachable((bw + dw, bh)):
push_loc = (bw + dw, bh)
hasspace = (self.is_loc_reachable((bw - dw, bh))
and self.is_loc_reachable((bw - 2 * dw, bh)))
if dw > 0:
push_direction = 'left'
else:
push_direction = 'right'
elif self.is_loc_reachable((bloc[0], bloc[1] + dh)):
push_loc = (bloc[0], bloc[1] + dh)
hasspace = (self.is_loc_reachable((bw, bh - dh))
and self.is_loc_reachable((bw, bh - 2 * dh)))
if dh > 0:
push_direction = 'down'
else:
push_direction = 'up'
else: #push loc is blocked, return with only 'stop'
return [[featurizer.featurize(self), 'stop']]
if not hasspace:
return [[featurizer.featurize(self), 'stop']]
p, bcost = dut.dijkstra_touch_cost(self, aloc, push_loc)
path = dut.collect_path(p, push_loc)
to_block = dut.path_to_actions(path)
actions.extend(to_block)
for a in to_block:
states.append(featurizer.featurize(self))
self.act(a)
self.update()
states.append(featurizer.featurize(self))
self.act('push_' + push_direction)
actions.append('push_' + push_direction)
self.update()
states.append(featurizer.featurize(self))
self.act(push_direction)
actions.append(push_direction)
self.update()
states.append(featurizer.featurize(self))
self.act('push_' + push_direction)
actions.append('push_' + push_direction)
self.update()
aloc = self.agent.attr['loc']
p, cost = dut.dijkstra_touch_cost(self, aloc, gloc)
if cost >= dut.get_big_cost():
# goal still unreachable (e.g. blocks next to door or pushed block onto goal)
# todo FIXME
return [[states[0], 'stop']]
path = dut.collect_path(p, gloc)
togoal = dut.path_to_actions(path)
actions.extend(togoal)
for a in togoal:
states.append(featurizer.featurize(self))
self.act(a)
self.update()
return list(zip(states, actions))
def get_supervision(self, featurizer, additional_featurizers=[]):
gloc = self.items_bytype['goal'][0].attr['loc']
path = []
all_actions = []
all_states = []
num_featurizers = len(additional_featurizers)
if num_featurizers > 0:
additional_rep = [[]] * num_featurizers
door = self.items_bytype['cycle_switch_opened_door'][0]
p, cost = dut.dijkstra_touch_cost(self, self.agent.attr['loc'], gloc)
if cost >= dut.get_big_cost():
# switch
# note: order issue when multiple switches exist
for switch in self.items_bytype['cycle_switch']:
switch_loc = switch.attr['loc']
p_switch, cost_switch = dut.dijkstra_touch_cost(
self, self.agent.attr['loc'], switch_loc)
if cost_switch >= dut.get_big_cost():
if num_featurizers == 0:
return [[featurizer.featurize(self), 'stop']]
else:
feat_list = [
f.featurize(self) for f in additional_featurizers
]
return [[
featurizer.featurize(self), 'stop', *feat_list
]]
path = dut.collect_path(p_switch, switch_loc)
actions = dut.path_to_actions(path)
for a in actions:
all_states.append(featurizer.featurize(self))
for i_f in range(num_featurizers):
additional_rep[i_f].append(
additional_featurizers[i_f].featurize(self))
self.act(a)
self.update()
all_actions += actions
while switch.color != door.color:
all_states.append(featurizer.featurize(self))
for i_f in range(num_featurizers):
additional_rep[i_f].append(
additional_featurizers[i_f].featurize(self))
all_actions.append('toggle_close')
self.act('toggle_close')
self.update()
# go to destination after toggling all switches to the same color as door
p, cost = dut.dijkstra_touch_cost(self, self.agent.attr['loc'], gloc)
path = dut.collect_path(p, gloc)
actions = dut.path_to_actions(path)
for a in actions:
all_states.append(featurizer.featurize(self))
for i_f in range(num_featurizers):
additional_rep[i_f].append(
additional_featurizers[i_f].featurize(self))
self.act(a)
self.update()
all_actions += actions
if num_featurizers == 0:
return list(zip(all_states, all_actions))
else:
return list(zip(all_states, all_actions, *additional_rep))