def pddl(self, costs, durative=False):
if durative:
return self.durative_pddl()
parameters_pddl = ' '.join(param.typed_pddl()
for param in self.parameters)
augmented_effect = self.effect
if costs and (
self.cost is not None
) and not self.cost_included: # Can only have one cost effect (uses last effect parsed)
augmented_effect = And(augmented_effect, Cost(self.cost))
return OPERATOR_PDDL.format(self._pddl_name, self.name,
parameters_pddl, self.condition.pddl(),
augmented_effect.pddl())
def __init__(self, oracle):
params = (P('q1', CONF), P('q2', CONF))
q1, q2 = params
super(Move, self).__init__(
self.__class__.__name__,
params,
[
AtConfig(q1),
#MoveCost(q1, q2),
#Initialize(MoveCost(q1, q2)), # TODO - should I define this here?
IsCollisionFree(q1, q2),
],
[
AtConfig(q2),
Not(AtConfig(q1)),
Cost(MoveCost(q1, q2)),
])
def get_stream_functions(universe):
action_to_function = {}
for action in universe.name_to_action.values():
if get_cost_atoms(action):
continue
if any(atom.predicate in universe.stream_predicates
for atom in action.condition.get_atoms()):
# TODO - assert that the atom is used positively
# TODO - alternatively, could make a stream action that has achieves the cost and plan with it
action_to_function[action] = Function(
FUNCTION_TEMPLATE % action.name,
[param.type for param in action.parameters])
universe.add_function(action_to_function[action])
function = action_to_function[action](*action.parameters)
action.effect = And(action.effect, Cost(function))
action.cost_included = True
return action_to_function
def compile_problem(estimator, task):
# Data types
CONF = Type()
SURFACE = Type() # Difference between fixed and movable objects
ITEM = Type()
POSE = Type()
CLASS = Type()
# Fluent predicates
AtConf = Pred(CONF)
HandEmpty = Pred()
Holding = Pred(ITEM)
AtPose = Pred(ITEM, POSE)
Supported = Pred(POSE, SURFACE) # Fluent
Localized = Pred(OBJECT)
Measured = Pred(OBJECT)
# Static predicates
IsKin = Pred(POSE, CONF)
IsClass = Pred(OBJECT, CLASS)
IsVisible = Pred(SURFACE, CONF)
IsSupported = Pred(POSE, SURFACE) # Static
# Functions
ScanRoom = Func(SURFACE)
#ScanTable = Func(SURFACE, TYPE)
ScanTable = Func(
SURFACE, ITEM) # TODO: could include more specific vantage point costs
Distance = Func(CONF, CONF)
# Derived
On = Pred(ITEM, SURFACE)
ComputableP = Pred(POSE)
ComputableQ = Pred(CONF)
# Free parameters
Q1, Q2 = Param(CONF), Param(CONF)
S1 = Param(SURFACE)
B1, B2 = Param(ITEM), Param(ITEM)
P1, P2 = Param(POSE), Param(POSE)
rename_easy(locals()) # Trick to make debugging easier
# TODO: could just do easier version of this that doesn't require localized to start
actions = [
Action(
name='pick',
parameters=[B1, P1, Q1], # TODO: Visibility constraint
condition=And(Localized(B1), AtPose(B1, P1), HandEmpty(),
AtConf(Q1), IsKin(P1, Q1)),
effect=And(Holding(B1), Not(AtPose(B1, P1)), Not(HandEmpty()))),
Action(name='place',
parameters=[B1, P1, Q1],
condition=And(Holding(B1), AtConf(Q1), IsKin(P1, Q1)),
effect=And(AtPose(B1, P1), HandEmpty(), Not(Holding(B1)))),
Action(name='move',
parameters=[Q1, Q2],
condition=And(AtConf(Q1), ComputableQ(Q2)),
effect=And(AtConf(Q2), Not(AtConf(Q1)), Cost(Distance(Q1,
Q2)))),
Action(name='scan_room',
parameters=[S1],
condition=Not(Localized(S1)),
effect=And(Localized(S1), Cost(ScanRoom(S1)))),
# TODO: need to set later poses to be usable or not to constrain order
Action(name='scan_table',
parameters=[S1, B1, P1, Q1],
condition=And(Localized(S1), AtConf(Q1), IsVisible(S1, Q1),
Not(Localized(B1))),
effect=And(Localized(B1), Measured(P1), Supported(P1, S1),
Cost(ScanTable(S1, B1)))),
]
axioms = [
# TODO: axiom for on? Might need a stream that generates initial fluents for On
# TODO: axiom that says that all fake values depending on a certain one now are usable
# TODO: could use stream predicates as fluents (as long as it doesn't break anything...)
Axiom(effect=On(B1, S1),
condition=Exists([P1],
And(AtPose(B1, P1),
Or(IsSupported(P1, S1), Supported(P1,
S1))))),
# TODO: compile automatically
Axiom(effect=ComputableQ(Q1),
condition=Or(Measured(Q1),
Exists([P1], And(IsKin(P1, Q1), ComputableP(P1))),
Exists([S1], And(IsVisible(S1, Q1),
Localized(S1))))),
Axiom(effect=ComputableP(P1),
condition=Or(
Measured(P1),
Exists([S1], And(IsSupported(P1, S1), Localized(S1))))),
]
#####
surface_types = estimator.surface_octomaps.keys()
item_types = estimator.object_octomaps.keys()
names_from_type = defaultdict(list)
known_poses = {}
holding = None
def add_type(cl):
name = '{}{}'.format(cl, len(names_from_type[cl]))
names_from_type[cl].append(name)
return name
# TODO: this is all very similar to the generic open world stuff
if estimator.holding is not None:
holding = add_type(estimator.holding)
for cl, octomap in estimator.surface_octomaps.items(
): # TODO: generic surface object
for pose in octomap.get_occupied():
known_poses[add_type(cl)] = pose
for cl, octomap in estimator.object_octomaps.items():
for pose in octomap.get_occupied():
known_poses[add_type(cl)] = pose
print dict(names_from_type), known_poses
# Human tells you to move block -> at least one block
# At least one block -> at least one surface
# TODO: generate fake properties about these fake values?
goal_objects, goal_surfaces = entities_from_task(task)
for cl in surface_types:
add_type(cl)
#for cl in goal_surfaces:
# for i in xrange(len(names_from_type[cl]), goal_surfaces[cl]):
# add_type(cl)
#for cl in item_types:
for cl in goal_objects:
for i in xrange(len(names_from_type[cl]), goal_objects[cl]):
add_type(cl)
#####
initial_atoms = [
AtConf(estimator.robot_conf),
Measured(CONF(estimator.robot_conf))
]
if holding is None:
initial_atoms.append(HandEmpty())
else:
initial_atoms.append(Holding(holding))
class_from_name = {
name: ty
for ty in names_from_type for name in names_from_type[ty]
}
for name, ty in class_from_name.iteritems():
ENTITY = SURFACE if ty in surface_types else ITEM
initial_atoms.append(IsClass(ENTITY(name), ty))
if name in known_poses:
initial_atoms.append(Localized(ENTITY(name)))
if ENTITY == ITEM:
pose = known_poses[name]
initial_atoms += [AtPose(name, pose), Measured(POSE(pose))]
else:
if ENTITY == ITEM:
pose = 'p_init_{}'.format(
name
) # The object should always be at this pose (we just can't do anything about it yet)
initial_atoms += [AtPose(name, pose)]
goal_literals = []
if task.holding is None:
goal_literals.append(HandEmpty())
elif task.holding is not False:
goal_literals.append(
Exists([B1], And(Holding(B1), IsClass(B1, task.holding))))
for obj, surface in task.object_surfaces:
goal_literals.append(
Exists([B1, S1],
And(On(B1, S1), IsClass(B1, obj), IsClass(S1, surface))))
goal_formula = And(*goal_literals)
####################
TOLERANCE = 0.1
def is_visible(table, conf):
x, y = conf
pose = known_poses[table]
#return (pose == x) and (y == 2)
#return (pose == x) and (y == 2)
return (pose == x) and (abs(y - 2) < TOLERANCE)
def is_kinematic(pose, conf):
x, y = conf
#return (pose == x) and (y == 1)
return (pose == x) and (abs(y - 1) < TOLERANCE)
####################
def sample_visible(table): # TODO: could generically do this with poses
if table in known_poses:
y = 2
#y += round(uniform(-TOLERANCE, TOLERANCE), 3)
conf = (known_poses[table], y)
assert is_visible(table, conf)
else:
conf = 'q_vis_{}'.format(table)
yield (conf, )
def inverse_kinematics(pose): # TODO: list stream that uses ending info
# TODO: only do if localized as well?
# TODO: is it helpful to have this even if the raw value is kind of wrong (to steer the search)
if type(pose) != str:
y = 1
#y += round(uniform(-TOLERANCE, TOLERANCE), 3)
conf = (pose, y)
assert is_kinematic(pose, conf)
else:
conf = 'q_ik_{}'.format(pose)
yield (conf, )
def sample_table(table):
if table in known_poses:
pose = known_poses[table]
else:
pose = 'p_{}'.format(table)
yield (pose, )
####################
MAX_DISTANCE = 10
# TODO: maybe I don't need to worry about normalizing. I can just pretend non-parametric again for planning
def scan_surface_cost(
surface, obj): # TODO: what about multiple scans of the belief?
fail_cost = 100
surface_cl = class_from_name[surface]
obj_cl = class_from_name[obj]
prob = 1.0
if obj_cl in estimator.object_prior:
prob *= estimator.object_prior[obj_cl].get(surface_cl, 0)
if surface in known_poses:
prob *= estimator.object_octomaps[obj_cl].get_prob(
known_poses[surface])
else:
prob *= 0.1 # Low chance if you don't even know the table exists
# TODO: could even include the probability the table exists
#return expected_cost(1, fail_cost, prob)
return mdp_cost(1, fail_cost, prob)
def scan_room_cost(surface):
# TODO: try to prove some sort of bound on the cost to recover will suffice?
fail_cost = 100
cl = class_from_name[surface]
occupied_poses = {
known_poses[n]
for n in names_from_type[cl] if n in known_poses
}
p_failure = 1.0
for pose in estimator.poses:
if pose not in occupied_poses:
p_failure *= (1 -
estimator.surface_octomaps[cl].get_prob(pose))
return 1 * (1 - p_failure) + fail_cost * p_failure
def distance_cost(q1, q2):
if str in (type(q1), type(q2)):
return MAX_DISTANCE # TODO: take the max possible pose distance
# TODO: can use the info encoded within these to obtain better bounds
return np.linalg.norm(np.array(q2) - np.array(q1))
####################
# TODO: could add measured as the output to these
streams = [
GeneratorStream(inputs=[P1],
outputs=[Q1],
conditions=[],
effects=[IsKin(P1, Q1)],
generator=inverse_kinematics),
GeneratorStream(inputs=[S1],
outputs=[Q1],
conditions=[],
effects=[IsVisible(S1, Q1)],
generator=sample_visible),
GeneratorStream(inputs=[S1],
outputs=[P1],
conditions=[],
effects=[IsSupported(P1, S1)],
generator=sample_table),
CostStream(inputs=[S1, B1],
conditions=[],
effects=[ScanTable(S1, B1)],
function=scan_surface_cost),
CostStream(inputs=[Q1, Q2],
conditions=[],
effects=[Distance(Q1, Q2),
Distance(Q2, Q1)],
function=distance_cost),
CostStream(inputs=[S1],
conditions=[],
effects=[ScanRoom(S1)],
function=scan_room_cost),
# TODO: make an is original precondition and only apply these to original values?
# I suppose I could apply to all concrete things but that's likely not useful
#TestStream(inputs=[S1, Q1], conditions=[IsOriginal(Q1)], effects=[IsVisible(S1, Q1)],
# test=is_visible, eager=True),
#TestStream(inputs=[P1, Q1], conditions=[IsOriginal(Q1), IsOriginal(Q1)], effects=[IsKin(P1, Q1)],
# test=is_kinematic, eager=True),
#GeneratorStream(inputs=[P1], outputs=[Q1], conditions=[], effects=[IsVisible(P1, Q1)],
# generator=sample_visible),
]
problem = STRIPStreamProblem(initial_atoms, goal_formula, actions + axioms,
streams, [])
def command_from_action((action, args)):
if action.name == 'scan_room':
return simulate_scan, []
if action.name in ('scan_table', 'look_block'):
return simulate_look, []
if action.name == 'move':
q1, q2 = map(get_value, args)
return simulate_move, [q2]
if action.name == 'pick':
o, p, q = map(get_value, args)
return simulate_pick, [class_from_name[o]]
if action.name == 'place':
return simulate_place, []
raise ValueError(action.name)
return problem, command_from_action
O, L1, L2, S = Param(OBJ), Param(LOC), Param(LOC), Param(STATE)
rename_easy(locals())
#UNKNOWN_LOC = 'unknown'
# NOTE - I could validate safe by inspecting all objects or just the goal location
# TODO - could do costs by either specifying them upfront or using STRIPStream
actions = [
Action(name='find', parameters=[O, L1], # Policy to find the object or an inspection?
#condition=And(At(O, UNKNOWN_LOC)),
#effect=And(At(O, L1), Not(At(O, UNKNOWN_LOC)))),
condition=And(UnsureLoc(O), Not(NotAtLoc(O, L1))),
effect=And(At(O, L1), Not(Clear(L1)), Not(UnsureLoc(O)), Cost(FindCost(O, L1)))), # TODO - can even adjust the cost of these to bias search towards more likely
Action(name='inspect_loc', parameters=[O, L1], # NOTE - O is not needed except for the execution
condition=And(UnsureClear(L1)),
effect=And(Clear(L1), Not(UnsureClear(L1)), Cost(InspectLocCost(L1)))),
Action(name='inspect_state', parameters=[O, L1, S],
condition=And(UnsureState(O), At(O, L1)), # Probably should know where it is before you worry about its state
#condition=And(UnsureState(O), Not(UnsureLoc(O))), # Probably should know where it is before you worry about its state
effect=And(HasState(O, S), Not(UnsureState(O)), Cost(InspectStateCost(O, S)))), # TODO - should I factor loc in the cost?
Action(name='transport', parameters=[O, L1, L2], # NOTE - Leslie and Tomas call this Move
#condition=And(At(O, L1)),
condition=And(At(O, L1), Clear(L2)),
effect=And(At(O, L2), Clear(L1), Not(At(O, L1)), Not(Clear(L2))),
cost=COST_SCALE*1),
##################################################
# NOTE - the location here is like a control parameter
actions = [
Action(
name='transport',
parameters=[O, L, L2, B, B2], # NOTE - Leslie and Tomas call this Move
condition=And(BAt(O, B), At(O, L), IsMoveUpdate(L, B, L2, B2)),
effect=And(
BAt(O, B2),
At(O, L2),
Not(At(O, L)),
Not(BAt(O, B)), #)),
Cost(MoveCost(L, B)))),
#Action(name='transport', parameters=[O, B, L, B2], # NOTE - I could include belief preconditions for safety or effectiveness
# condition=And(BAt(O, B), BAtAbove(L, B2, ), IsMoveUpdate(B, L, B2)),
# effect=And(BAt(O, B2), Not(BAt(O, B)))), # NOTE - Leslie and Tomas call this Move
Action(
name='find',
parameters=[O, L, B, B2],
condition=And(BAt(O, B), At(O, L), IsLookUpdate(B, L, B2)),
effect=And(
BAt(O, B2),
Not(BAt(O, B)), #)),
Cost(LookCost(B, L)))),
Action(
name='infer_at',
parameters=[O, L],