def test_hierarchical_spaces():
dir_path = os.path.dirname(os.path.realpath(__file__))
domain_file = os.path.join(dir_path, 'pddl',
'hierarchical_type_test_domain.pddl')
problem_file = os.path.join(dir_path, 'pddl',
'hierarchical_type_test_domain',
'hierarchical_type_test_problem.pddl')
domain = PDDLDomainParser(domain_file)
problem = PDDLProblemParser(problem_file, domain.domain_name, domain.types,
domain.predicates, domain.actions)
actions = list(domain.actions)
action_predicates = [domain.predicates[a] for a in actions]
space = LiteralSpace(set(domain.predicates.values()) -
set(action_predicates),
type_to_parent_types=domain.type_to_parent_types)
all_ground_literals = space.all_ground_literals(
State(problem.initial_state, problem.objects, problem.goal))
ispresent = Predicate("ispresent", 1, [Type("entity")])
islight = Predicate("islight", 1, [Type("object")])
isfurry = Predicate("isfurry", 1, [Type("animal")])
ishappy = Predicate("ishappy", 1, [Type("animal")])
attending = Predicate("attending", 2, [Type("animal"), Type("object")])
nomsy = Type("jindo")("nomsy")
rover = Type("corgi")("rover")
rene = Type("cat")("rene")
block1 = Type("block")("block1")
block2 = Type("block")("block2")
cylinder1 = Type("cylinder")("cylinder1")
assert all_ground_literals == {
ispresent(nomsy),
ispresent(rover),
ispresent(rene),
ispresent(block1),
ispresent(block2),
ispresent(cylinder1),
islight(block1),
islight(block2),
islight(cylinder1),
isfurry(nomsy),
isfurry(rover),
isfurry(rene),
ishappy(nomsy),
ishappy(rover),
ishappy(rene),
attending(nomsy, block1),
attending(nomsy, block2),
attending(nomsy, cylinder1),
attending(rover, block1),
attending(rover, block2),
attending(rover, cylinder1),
attending(rene, block1),
attending(rene, block2),
attending(rene, cylinder1),
}
print("Test passed.")
def _parse_domain_predicates(self):
start_ind = re.search(r"\(:predicates", self.domain).start()
predicates = self._find_balanced_expression(self.domain, start_ind)
predicates = predicates[12:-1].strip()
predicates = self._find_all_balanced_expressions(predicates)
self.predicates = {}
for pred in predicates:
pred = pred.strip()[1:-1].split("?")
pred_name = pred[0].strip()
# arg_types = [self.types[arg.strip().split("-")[1].strip()]
# for arg in pred[1:]]
arg_types = []
for arg in pred[1:]:
if ' - ' in arg:
assert arg_types is not None, "Mixing of typed and untyped args not allowed"
assert self.uses_typing
arg_type = self.types[arg.strip().split("-", 1)[1].strip()]
arg_types.append(arg_type)
else:
assert not self.uses_typing
arg_types.append(self.types["default"])
self.predicates[pred_name] = Predicate(pred_name, len(pred[1:]),
arg_types)
# Handle equality
if "=" in self.domain:
self.predicates["="] = Predicate("=", 2)
def test_zero_arity_negative_preconditions():
MoveableType = Type('moveable')
Holding = Predicate('Holding', 1, var_types=[MoveableType])
HandEmpty = Predicate('HandEmpty', 0, var_types=[])
conds = [Holding("?x1"), Not(HandEmpty())]
kb = {Holding("a"), HandEmpty()}
assignments = find_satisfying_assignments(kb,
conds,
allow_redundant_variables=False)
assert len(assignments) == 0
print("Pass.")
def test_heuristic():
dir_path = os.path.dirname(os.path.realpath(__file__))
domain_file = os.path.join(dir_path, 'pddl', 'easyblocks.pddl')
problem_dir = os.path.join(dir_path, 'pddl', 'easyblocks')
block = Type("block")
pickup = Predicate("pickup", 1, [block])
stack = Predicate("stack", 2, [block, block])
shaped_env = PDDLEnv(
domain_file,
problem_dir,
shape_reward_mode="lmcut",
raise_error_on_invalid_action=True,
)
env = PDDLEnv(
domain_file,
problem_dir,
shape_reward_mode=None,
raise_error_on_invalid_action=True,
)
shaped_env.reset()
env.reset()
_, rew, _, _ = env.step(pickup("b"))
_, shaped_rew, _, _ = shaped_env.step(pickup("b"))
assert rew == 0.
assert shaped_rew == 1.
intermediate_state = env.get_state()
def assert_last_step():
_, rew, done, _ = env.step(stack("b", "a"))
_, shaped_rew, shaped_done, _ = shaped_env.step(stack("b", "a"))
assert done
assert shaped_done
assert rew == 1.
assert shaped_rew == 2.
# check if the step to the goal is terminal and with correct rewards
assert_last_step()
# check if shaped reward is consistent after setting the state
shaped_env.set_state(intermediate_state)
env.set_state(intermediate_state)
assert_last_step()
print("Test passed.")
def effect_to_literal(literal):
assert isinstance(literal, Literal)
assert literal.predicate.name.startswith("Effect")
non_effect_pred = Predicate(literal.predicate.name[len("Effect"):],
literal.predicate.arity,
literal.predicate.var_types,
is_negative=literal.predicate.is_negative,
is_anti=literal.predicate.is_anti)
return non_effect_pred(*literal.variables)
def _create_actions_from_operators(self):
actions = set()
for name, operator in self.operators.items():
types = [p.var_type for p in operator.params]
action = Predicate(name, len(types), types)
assert name not in self.predicates, "Cannot have predicate with same name as operator"
self.predicates[name] = action
actions.add(action)
return actions
def _preprocess_negative_literals(cls, kb, conds):
# Check for negated quantifiers, which we do not handle
if any((isinstance(c, Exists) or isinstance(c, ForAll)) and c.is_negative \
for c in conds):
raise NotImplementedError("We do not yet handle negated quantifiers")
# Find all predicates with a negated literal in the conds
negated_predicates = set()
for cond in cls._get_lits_from_conds(conds):
if cond.is_negative:
negated_predicates.add(cond.predicate)
if len(negated_predicates) == 0:
return kb, conds
# Start the new kb and conds
kb = [lit for lit in kb]
conds = [c for c in conds]
# Sanity check
assert all(str(p).startswith("Not") for p in negated_predicates)
# Create positive predicates for the negated predicates
negated_pred_to_pos_pred = {}
for p in negated_predicates:
# Prolog hands = specially
if p.name == "=":
pos_pred = Predicate(f"neg-eq", p.arity, p.var_types)
else:
pos_pred = Predicate(f"neg-{p.name}", p.arity, p.var_types)
negated_pred_to_pos_pred[p] = pos_pred
# TODO pass in objects separately
objects = { o for lit in kb for o in lit.variables }
# Get all instantiations of the new positive predicates
for negated_pred, pos_pred in negated_pred_to_pos_pred.items():
original_positive_pred = negated_pred.positive
# Get all combinations of objects
for objs in get_object_combinations(objects,
arity=pos_pred.arity,
var_types=pos_pred.var_types,
allow_duplicates=True):
# Check whether the positive version is in the kb
if original_positive_pred(*objs) in kb:
continue
# Add the new positive literal to the kb
kb.append(pos_pred(*objs))
# Update the conds to include the positive pred instead of the negative pred
conds = cls._replace_predicate(conds, negated_pred, pos_pred)
return kb, conds
def wrap_goal_literal(cls, x):
"""Helper for converting a state to required input representation
"""
if isinstance(x, Predicate):
return Predicate("WANT" + x.name,
x.arity,
var_types=x.var_types,
is_negative=x.is_negative,
is_anti=x.is_anti)
new_predicate = cls.wrap_goal_literal(x.predicate)
return new_predicate(*x.variables)
def test_negative_preconditions():
MoveableType = Type('moveable')
Holding = Predicate('Holding', 1, var_types=[MoveableType])
IsPawn = Predicate('IsPawn', 1, var_types=[MoveableType])
PutOn = Predicate('PutOn', 1, var_types=[MoveableType])
On = Predicate('On', 2, var_types=[MoveableType, MoveableType])
# ?x0 must bind to o0 and ?x1 must bind to o1, so ?x2 must bind to o2
conds = [
PutOn("?x0"),
Holding("?x1"),
IsPawn("?x2"),
Not(On("?x2", "?x0"))
]
kb = {
PutOn('o0'),
IsPawn('o0'),
IsPawn('o1'),
IsPawn('o2'),
Holding('o1'),
}
assignments = find_satisfying_assignments(kb,
conds,
allow_redundant_variables=False)
assert len(assignments) == 1
# should be the same, even though IsPawn("?x2") is removed
conds = [PutOn("?x0"), Holding("?x1"), Not(On("?x2", "?x0"))]
kb = {
PutOn('o0'),
IsPawn('o0'),
IsPawn('o1'),
IsPawn('o2'),
Holding('o1'),
}
assignments = find_satisfying_assignments(kb,
conds,
allow_redundant_variables=False)
assert len(assignments) == 1
print("Pass.")
def Effect(x): # pylint:disable=invalid-name
"""An effect predicate or literal.
"""
if isinstance(x, Predicate):
return Predicate("Effect" + x.name,
x.arity,
var_types=x.var_types,
is_negative=x.is_negative,
is_anti=x.is_anti)
assert isinstance(x, Literal)
new_predicate = Effect(x.predicate)
return new_predicate(*x.variables)
def reverse_binary_literal(cls, x):
"""Helper for converting a state to required input representation
"""
if isinstance(x, Predicate):
assert x.arity == 2
return Predicate("REV" + x.name,
x.arity,
var_types=x.var_types,
is_negative=x.is_negative,
is_anti=x.is_anti)
new_predicate = cls.reverse_binary_literal(x.predicate)
variables = [v for v in x.variables]
assert len(variables) == 2
return new_predicate(*variables[::-1])
def test_determinize():
dir_path = os.path.dirname(os.path.realpath(__file__))
domain_file = os.path.join(dir_path, 'pddl',
'test_probabilistic_domain.pddl')
problem_dir = os.path.join(dir_path, 'pddl', 'test_probabilistic_domain')
env = PDDLEnv(domain_file,
problem_dir,
raise_error_on_invalid_action=True,
dynamic_action_space=True)
env.domain.determinize()
obs, _ = env.reset()
action = env.action_space.all_ground_literals(obs).pop()
transitions = env.get_all_possible_transitions(action, return_probs=True)
transition_list = list(transitions)
assert len(transition_list) == 1
assert transition_list[0][1] == 1.0
newstate = transition_list[0][0][0]
type1 = Type('type1')
type2 = Type('type2')
pred1 = Predicate('pred1', 1, [type1])
pred2 = Predicate('pred2', 1, [type2])
pred3 = Predicate('pred3', 3, [type1, type2, type2])
assert newstate.literals == frozenset({
pred1('b2'),
pred2('c1'),
pred3('a1', 'c1', 'd1'),
pred3('a2', 'c2', 'd2'),
pred3('b2', 'd1', 'c1')
})
print("Test passed.")
def _create_problem_file(self, raw_problem_fname, use_cache=True):
if (not use_cache) or (raw_problem_fname not in self._problem_files):
problem_fname = os.path.split(raw_problem_fname)[-1]
problem_fname = problem_fname.split('.pddl')[0]
problem_fname += '_{}_with_diffs_{}.pddl'.format(
self.domain_name, random.randint(0, 9999999))
problem_fname = os.path.join('/tmp', problem_fname)
# Parse raw problem
problem_parser = PDDLProblemParser(raw_problem_fname,
self.domain_name.lower(),
self._types, self._predicates,
self._actions)
new_initial_state = set(problem_parser.initial_state)
# Add actions
action_lits = set(
self._action_space.all_ground_literals(State(
new_initial_state, problem_parser.objects,
problem_parser.goal),
valid_only=False))
new_initial_state |= action_lits
# Add 'Different' pairs for each pair of objects
Different = Predicate('Different', 2)
for obj1 in problem_parser.objects:
for obj2 in problem_parser.objects:
if obj1 == obj2:
continue
# if obj1.var_type != obj2.var_type:
# continue
diff_lit = Different(obj1, obj2)
new_initial_state.add(diff_lit)
# Write out new temporary problem file
problem_parser.initial_state = frozenset(new_initial_state)
problem_parser.write(problem_fname)
# Add to cache
self._problem_files[raw_problem_fname] = problem_fname
return self._problem_files[raw_problem_fname]
def test_pddlenv():
dir_path = os.path.dirname(os.path.realpath(__file__))
domain_file = os.path.join(dir_path, 'pddl', 'test_domain.pddl')
problem_dir = os.path.join(dir_path, 'pddl', 'test_domain')
env = PDDLEnv(domain_file,
problem_dir,
raise_error_on_invalid_action=True,
dynamic_action_space=True)
env2 = PDDLEnv(domain_file,
problem_dir,
raise_error_on_invalid_action=True,
dynamic_action_space=False)
type1 = Type('type1')
type2 = Type('type2')
pred1 = Predicate('pred1', 1, [type1])
pred2 = Predicate('pred2', 1, [type2])
pred3 = Predicate('pred3', 1, [type1, type2, type2])
operator_name = 'action1'
action_pred = Predicate('actionpred', 1, [type1])
obs, _ = env.reset()
assert obs == {
pred1('b2'),
pred2('c1'),
pred3('a1', 'c1', 'd1'),
pred3('a2', 'c2', 'd2')
}
operator = env.domain.operators[operator_name]
# Invalid action
action = action_pred('b1')
try:
env.step(action)
assert False, "Action was supposed to be invalid"
except InvalidAction:
pass
assert action not in env.action_space.all_ground_literals(
), "Dynamic action space not working"
env2.reset()
assert action in env2.action_space.all_ground_literals(
), "Dynamic action space not working"
# Valid args
action = action_pred('b2')
obs, _, _, _ = env.step(action)
assert obs == {
pred1('b2'),
pred2('c1'),
pred2('c2'),
pred3('b2', 'd2', 'c2'),
pred3('a1', 'c1', 'd1'),
pred3('a2', 'c2', 'd2')
}
print("Test passed.")
def test_pddlenv_hierarchical_types():
dir_path = os.path.dirname(os.path.realpath(__file__))
domain_file = os.path.join(dir_path, 'pddl',
'hierarchical_type_test_domain.pddl')
problem_dir = os.path.join(dir_path, 'pddl',
'hierarchical_type_test_domain')
env = PDDLEnv(domain_file, problem_dir)
obs, _ = env.reset()
ispresent = Predicate("ispresent", 1, [Type("entity")])
islight = Predicate("islight", 1, [Type("object")])
isfurry = Predicate("isfurry", 1, [Type("animal")])
ishappy = Predicate("ishappy", 1, [Type("animal")])
pet = Predicate("pet", 1, [Type("animal")])
nomsy = Type("jindo")("nomsy")
rover = Type("corgi")("rover")
rene = Type("cat")("rene")
block1 = Type("block")("block1")
block2 = Type("block")("block2")
cylinder1 = Type("cylinder")("cylinder1")
assert obs.literals == frozenset({
ispresent(nomsy),
ispresent(rover),
ispresent(rene),
ispresent(block1),
ispresent(block2),
ispresent(cylinder1),
islight(block1),
islight(cylinder1),
isfurry(nomsy),
})
obs, _, _, _ = env.step(pet('block1'))
assert obs.literals == frozenset({
ispresent(nomsy),
ispresent(rover),
ispresent(rene),
ispresent(block1),
ispresent(block2),
ispresent(cylinder1),
islight(block1),
islight(cylinder1),
isfurry(nomsy),
})
obs, _, _, _ = env.step(pet(nomsy))
assert obs.literals == frozenset({
ispresent(nomsy),
ispresent(rover),
ispresent(rene),
ispresent(block1),
ispresent(block2),
ispresent(cylinder1),
islight(block1),
islight(cylinder1),
isfurry(nomsy),
ishappy(nomsy)
})
print("Test passed.")
def test_get_all_possible_transitions_multiple_effects():
dir_path = os.path.dirname(os.path.realpath(__file__))
domain_file = os.path.join(dir_path, 'pddl',
'test_probabilistic_domain_alt_2.pddl')
problem_dir = os.path.join(dir_path, 'pddl',
'test_probabilistic_domain_alt_2')
env = PDDLEnv(domain_file,
problem_dir,
raise_error_on_invalid_action=True,
dynamic_action_space=True)
obs, _ = env.reset()
action = env.action_space.all_ground_literals(obs).pop()
transitions = env.get_all_possible_transitions(action)
transition_list = list(transitions)
assert len(transition_list) == 3
states = set({
transition_list[0][0].literals, transition_list[1][0].literals,
transition_list[2][0].literals
})
type1 = Type('type1')
type2 = Type('type2')
pred1 = Predicate('pred1', 1, [type1])
pred2 = Predicate('pred2', 1, [type2])
pred3 = Predicate('pred3', 3, [type1, type2, type2])
expected_states = set({
frozenset(
{pred1('b2'),
pred3('a1', 'c1', 'd1'),
pred3('a2', 'c2', 'd2')}),
frozenset(
{pred2('c1'),
pred3('a1', 'c1', 'd1'),
pred3('a2', 'c2', 'd2')}),
frozenset({
pred1('b2'),
pred2('c1'),
pred3('a1', 'c1', 'd1'),
pred3('a2', 'c2', 'd2'),
pred3('b2', 'd1', 'c1')
})
})
assert states == expected_states
# Now test again with return_probs=True.
transitions = env.get_all_possible_transitions(action, return_probs=True)
transition_list = list(transitions)
assert len(transition_list) == 3
states_and_probs = {
transition_list[0][0][0].literals: transition_list[0][1],
transition_list[1][0][0].literals: transition_list[1][1],
transition_list[2][0][0].literals: transition_list[2][1]
}
type1 = Type('type1')
type2 = Type('type2')
pred1 = Predicate('pred1', 1, [type1])
pred2 = Predicate('pred2', 1, [type2])
pred3 = Predicate('pred3', 3, [type1, type2, type2])
expected_states = {
frozenset({
pred1('b2'),
pred3('a1', 'c1', 'd1'),
pred3('a2', 'c2', 'd2')
}):
0.5,
frozenset({
pred2('c1'),
pred3('a1', 'c1', 'd1'),
pred3('a2', 'c2', 'd2')
}):
0.4,
frozenset({
pred1('b2'),
pred2('c1'),
pred3('a1', 'c1', 'd1'),
pred3('a2', 'c2', 'd2'),
pred3('b2', 'd1', 'c1')
}):
0.1
}
for s, prob in states_and_probs.items():
assert s in expected_states
assert prob - expected_states[s] < 1e-5
print("Test passed.")
class FOLDTOperatorLearningModule:
_NoChange = Predicate("NoChange", 0)
def __init__(self, learned_operators):
self._learned_operators = learned_operators
self._learned_operators_for_action = defaultdict(set)
self._Xs = defaultdict(list)
self._Ys = defaultdict(list)
self._seed = ac.seed
self._rand_state = np.random.RandomState(seed=ac.seed)
self.learned_dts = {}
self._fits_all_data = defaultdict(bool)
self._learning_on = True
def observe(self, state, action, effects):
if not self._learning_on:
return
x = (state.literals | {action})
y = sorted([Effect(e) for e in effects])
if not y:
y.append(Effect(self._NoChange()))
self._Xs[action.predicate].append(x)
self._Ys[action.predicate].append(y)
# print("observed data:")
# print(x)
# print(y)
# input("!!")
# Check whether we'll need to relearn
if self._fits_all_data[action.predicate]:
dt = self.learned_dts[action.predicate]
if not self._dt_fits_data(dt, x, y):
self._fits_all_data[action.predicate] = False
def learn(self):
if not self._learning_on:
return False
is_updated = False
for action_predicate in self._fits_all_data:
if not self._fits_all_data[action_predicate]:
# Learn one main tree.
try:
dt = self._learn_single_dt(action_predicate, self._seed,
self._Xs[action_predicate],
self._Ys[action_predicate])
except LearningFailure:
# Aggressive!!!!!!!!!
self._Xs[action_predicate] = []
self._Ys[action_predicate] = []
continue
self.learned_dts[action_predicate] = dt
operators = self._foldt_to_operators(dt, action_predicate.name)
for op in operators:
if not sum(lit.predicate.name == action_predicate
for lit in op.preconds.literals) == 1:
# Something is wrong...why are multiple action
# predicates appearing in the preconditions?
import ipdb
ipdb.set_trace()
# Remove old operators
self._learned_operators_for_action[action_predicate].clear()
# Add new operators
self._learned_operators_for_action[action_predicate].update(
operators)
self._fits_all_data[action_predicate] = True
is_updated = True
# Update all learned operators
if is_updated:
self._learned_operators.clear()
for operators in self._learned_operators_for_action.values():
self._learned_operators.update(operators)
return is_updated
@staticmethod
def _learn_single_dt(action_predicate,
seed,
X,
Y,
bag_predicates=False,
bag_features=False):
root_literal = action_predicate(
*
['Placeholder{}'.format(i) for i in range(action_predicate.arity)])
dt = FOLDTClassifier(
seed=seed,
root_feature=root_literal,
max_feature_length=ac.max_foldt_feature_length,
max_learning_time=ac.max_foldt_learning_time,
max_exceeded_strategy=ac.max_foldt_exceeded_strategy,
bag_predicates=bag_predicates,
bag_features=bag_features)
dt.fit(X, Y)
return dt
def turn_off(self):
self._learning_on = False
def _dt_fits_data(self, dt, x, y):
prediction = self.get_prediction(x, dt)
if prediction is None:
match = y is None
else:
match = set(y) == set(prediction)
# if not match:
# print("Mismatch:")
# print("x =", x)
# print("y =", y)
# print("prediction =", prediction)
return match
@classmethod
def get_prediction(cls, x, dt):
prediction = dt.predict(x)
if prediction is None:
return prediction
# Cancel out effects if possible (e.g. At(x) and AntiAt(x))
final_prediction = set()
for lit in prediction:
if lit.inverted_anti in final_prediction:
final_prediction.remove(lit.inverted_anti)
else:
final_prediction.add(lit)
if not final_prediction:
final_prediction = [Effect(cls._NoChange())]
return sorted(list(final_prediction))
def _foldt_to_operators(self, dt, action_name, suffix=''):
op_counter = 0
operators = set()
for (path, leaf) in dt.get_conditional_literals(dt.root):
if leaf is None:
continue
if any(lit.predicate == Effect(self._NoChange) for lit in leaf):
continue
name = "LearnedOperator{}{}{}".format(action_name, op_counter,
suffix)
op_counter += 1
# NoChange appears only in effects in the training data, so it
# should never be in the preconditions of the learned operators.
assert not any(lit.predicate.positive == self._NoChange
for lit in path)
preconds = LiteralConjunction(path)
effects = LiteralConjunction([effect_to_literal(l) for l in leaf])
params = self._get_params_from_preconds(preconds)
operator = Operator(name, params, preconds, effects)
operators.add(operator)
return operators
def _get_params_from_preconds(self, preconds):
param_set = set()
for lit in preconds.literals:
if lit.negated_as_failure:
continue
for v in lit.variables:
param_set.add(v)
return sorted(list(param_set))
def get_sar_successor_state(state, action):
"""Search and rescue specific successor generation
Assumptions:
- One robot called robot0
"""
# Remake predicates to keep this function self-contained
person_type = Type('person')
robot_type = Type('robot')
location_type = Type('location')
direction_type = Type('direction')
clear = Predicate('clear', 1, [location_type])
conn = Predicate("conn", 3, [location_type, location_type, direction_type])
robot_at = Predicate('robot-at', 2, [robot_type, location_type])
carrying = Predicate('carrying', 2, [robot_type, person_type])
person_at = Predicate('person-at', 2, [person_type, location_type])
handsfree = Predicate('handsfree', 1, [robot_type])
# Parse the state
robot_location = None # location
robot_carrying = None # person
adjacency_map = {} # (location, direction) -> location
people_locs = {} # person -> location
clear_locs = set()
for lit in state.literals:
if lit.predicate.name == "robot-at":
assert lit.variables[0] == "robot0"
robot_location = lit.variables[1]
elif lit.predicate.name == "conn":
adjacency_map[(lit.variables[0], lit.variables[2])] = lit.variables[1]
elif lit.predicate.name == "carrying":
assert lit.variables[0] == "robot0"
robot_carrying = lit.variables[1]
elif lit.predicate.name == "person-at":
people_locs[lit.variables[0]] = lit.variables[1]
elif lit.predicate.name == "clear":
clear_locs.add(lit.variables[0])
assert robot_location is not None
is_valid = False
pos_preconds = set()
neg_preconds = set()
pos_effects = set()
neg_effects = set()
if action.predicate.name == "move":
"""
(:action move-robot
:parameters (?robot - robot ?from - location ?to - location ?dir - direction)
:precondition (and (move ?dir)
(conn ?from ?to ?dir)
(robot-at ?robot ?from)
(clear ?to))
:effect (and
(not (robot-at ?robot ?from))
(robot-at ?robot ?to)
(not (clear ?to))
(clear ?from))
)
"""
direction = action.variables[0]
if (robot_location, direction) in adjacency_map:
next_robot_location = adjacency_map[(robot_location, direction)]
if next_robot_location in clear_locs:
is_valid = True
pos_preconds = {
conn(robot_location, next_robot_location, direction),
robot_at("robot0", robot_location),
clear(next_robot_location),
}
pos_effects = {
robot_at("robot0", next_robot_location),
clear(robot_location)
}
neg_effects = {
robot_at("robot0", robot_location),
clear(next_robot_location)
}
elif action.predicate.name == "pickup":
"""
(:action pickup-person
:parameters (?robot - robot ?person - person ?loc - location)
:precondition (and (pickup ?person)
(robot-at ?robot ?loc)
(person-at ?person ?loc)
(handsfree ?robot))
:effect (and
(not (person-at ?person ?loc))
(not (handsfree ?robot))
(carrying ?robot ?person))
)
"""
if robot_carrying is None:
person = action.variables[0]
person_loc = people_locs[person]
if person_loc == robot_location:
is_valid = True
pos_preconds = {
robot_at("robot0", robot_location),
person_at(person, person_loc),
handsfree("robot0"),
}
pos_effects = {
carrying("robot0", person),
}
neg_effects = {
person_at(person, person_loc),
handsfree("robot0"),
}
elif action.predicate.name == "dropoff":
"""
(:action dropoff-person
:parameters (?robot - robot ?person - person ?loc - location)
:precondition (and (dropoff )
(carrying ?robot ?person)
(robot-at ?robot ?loc))
:effect (and
(person-at ?person ?loc)
(handsfree ?robot)
(not (carrying ?robot ?person)))
)
"""
if robot_carrying is not None:
is_valid = True
pos_preconds = {
robot_at("robot0", robot_location),
carrying("robot0", robot_carrying),
}
pos_effects = {
person_at(robot_carrying, robot_location),
handsfree("robot0"),
}
neg_effects = {
carrying("robot0", robot_carrying),
}
else:
raise Exception(f"Unrecognized action {action}.")
if not is_valid:
return state
assert state.literals.issuperset(pos_preconds)
assert len(state.literals & {Not(p) for p in neg_preconds}) == 0
new_state_literals = set(state.literals)
new_state_literals -= neg_effects
new_state_literals |= pos_effects
return state.with_literals(frozenset(new_state_literals))
def test_prover():
TType = Type('t')
atom0, atom1, atom2 = TType('atom0'), TType('atom1'), TType('atom2')
var0, var1, var2, var3 = TType('Var0'), TType('Var1'), TType(
'Var2'), TType('Var3')
# Single predicate single arity test
predicate0 = Predicate('Predicate0', 1, [TType])
predicate1 = Predicate('Predicate1', 2, [TType, TType])
predicate2 = Predicate('Predicate2', 1, [TType])
kb0 = [predicate0(atom0)]
assignments = find_satisfying_assignments(kb0, [predicate0(var0)])
assert len(assignments) == 1
assert len(assignments[0]) == 1
assert assignments[0][var0] == atom0
assignments = find_satisfying_assignments(
kb0, [predicate0(var0), predicate0(var1)])
assert len(assignments) == 1
kb1 = [predicate0(atom0), predicate0(atom1)]
assignments = find_satisfying_assignments(kb1, [predicate0(var0)])
assert len(assignments) == 2
assignments = find_satisfying_assignments(
kb1, [predicate0(var0), predicate0(var1)])
assert len(assignments) == 2
assignments = find_satisfying_assignments(
kb1, [predicate0(var0),
predicate0(var1),
predicate0(var2)])
assert len(assignments) == 2
kb2 = [predicate0(atom0), predicate0(atom1), predicate0(atom2)]
assignments = find_satisfying_assignments(kb2, [predicate0(var0)])
assert len(assignments) == 2
assignments = find_satisfying_assignments(
kb2, [predicate0(var0), predicate0(var1)])
assert len(assignments) == 2
assignments = find_satisfying_assignments(
kb2, [predicate0(var0),
predicate0(var1),
predicate0(var2)])
assert len(assignments) == 2
# Single predicate multiple arity test
kb3 = [predicate1(atom0, atom1), predicate1(atom1, atom2)]
assignments = find_satisfying_assignments(kb3, [predicate1(var0, var1)])
assert len(assignments) == 2
assignments = find_satisfying_assignments(
kb3, [predicate1(var0, var1),
predicate1(var1, var2)])
assert len(assignments) == 1
assignments = find_satisfying_assignments(
kb3, [predicate1(var0, var1),
predicate1(var1, var0)])
assert len(assignments) == 0
assignments = find_satisfying_assignments(
kb3, [predicate1(var0, var1),
predicate1(var2, var3)])
assert len(assignments) == 2
## Multiple predicate multiple arity test
kb4 = [
predicate0(atom2),
predicate1(atom0, atom1),
predicate1(atom1, atom2)
]
assignments = find_satisfying_assignments(
kb4, [predicate1(var0, var1),
predicate0(var1),
predicate0(var0)])
assert len(assignments) == 0
## Tricky case!
kb6 = [
predicate0(atom0),
predicate2(atom1),
predicate1(atom0, atom2),
predicate1(atom2, atom1)
]
assignments = find_satisfying_assignments(
kb6, [predicate0(var0),
predicate2(var1),
predicate1(var0, var1)])
assert len(assignments) == 0
print("Pass.")
from pddlgym.structs import Predicate, ground_literal, LiteralConjunction
from pddlgym.parser import Operator
from pddlgym.inference import find_satisfying_assignments
import numpy as np
### Noisy deictic rules
NOISE_OUTCOME = Predicate("noiseoutcome", 0, [])()
class MultipleOutcomesPossible(Exception):
pass
class NDR:
"""A Noisy Deictic Rule has a lifted action, lifted preconditions, and a
distribution over effects.
Parameters
----------
action : Literal
preconditions : [ Literal ]
effect_probs : np.ndarray
effects : [Literal]
"""
def __init__(self,
action,
preconditions,
effect_probs,
effects,
allow_redundant_variables=False):
self._action = action
class SearchAndRescueEnv(PDDLSearchAndRescueEnv):
"""Changes the state space to just be positions of objects
and the identity of the person being carried.
"""
person_type = Type('person')
robot_type = Type('robot')
location_type = Type('location')
direction_type = Type('direction')
wall_type = Type('wall')
hospital_type = Type('hospital')
clear = Predicate('clear', 1, [location_type])
conn = Predicate("conn", 3, [location_type, location_type, direction_type])
robot_at = Predicate('robot-at', 2, [robot_type, location_type])
person_at = Predicate('person-at', 2, [person_type, location_type])
wall_at = Predicate('wall-at', 2, [wall_type, location_type])
hospital_at = Predicate('hospital-at', 2, [hospital_type, location_type])
carrying = Predicate('carrying', 2, [robot_type, person_type])
handsfree = Predicate('handsfree', 1, [robot_type])
@property
def observation_space(self):
raise NotImplementedError()
def _state_to_internal(self, state):
state = dict(state)
new_state_literals = set()
directions_to_deltas = {
self.direction_type('up') : (-1, 0),
self.direction_type('down') : (1, 0),
self.direction_type('left') : (0, -1),
self.direction_type('right') : (0, 1),
}
# conn
height, width = 6, 6
for r in range(height):
for c in range(width):
loc = self.location_type(f"f{r}-{c}f")
for direction, (dr, dc) in directions_to_deltas.items():
next_r, next_c = r + dr, c + dc
if not (0 <= next_r < height and 0 <= next_c < width):
continue
next_loc = self.location_type(f"f{next_r}-{next_c}f")
conn_lit = self.conn(loc, next_loc, direction)
new_state_literals.add(conn_lit)
# at
occupied_locs = set()
hospital_loc = None
for obj_name, loc_tup in state.items():
if obj_name in ["carrying", "rescue"]:
continue
r, c = loc_tup
loc = self.location_type(f"f{r}-{c}f")
if obj_name.startswith("person"):
at_pred = self.person_at
elif obj_name.startswith("robot"):
at_pred = self.robot_at
occupied_locs.add((r, c))
elif obj_name.startswith("wall"):
at_pred = self.wall_at
occupied_locs.add((r, c))
elif obj_name.startswith("hospital"):
at_pred = self.hospital_at
assert hospital_loc is None
hospital_loc = loc
else:
raise Exception(f"Unrecognized object {obj_name}")
new_state_literals.add(at_pred(obj_name, loc))
assert hospital_loc is not None
# carrying/handsfree
if state["carrying"] is None:
new_state_literals.add(self.handsfree("robot0"))
else:
new_state_literals.add(self.carrying("robot0", state["carrying"]))
# clear
for r in range(height):
for c in range(width):
if (r, c) not in occupied_locs:
loc = self.location_type(f"f{r}-{c}f")
clear_lit = self.clear(loc)
new_state_literals.add(clear_lit)
# objects
new_objects = frozenset({o for lit in new_state_literals for o in lit.variables })
# goal
new_goal = LiteralConjunction([self.person_at(person, hospital_loc) \
for person in sorted(state["rescue"])])
new_state = State(frozenset(new_state_literals), new_objects, new_goal)
return new_state
def _internal_to_state(self, internal_state):
state = { "carrying" : None }
state["rescue"] = set()
for lit in internal_state.goal.literals:
assert lit.predicate == self.person_at
state["rescue"].add(lit.variables[0].name)
state["rescue"] = frozenset(state["rescue"]) # make hashable
for lit in internal_state.literals:
if lit.predicate.name.endswith("at"):
obj_name = lit.variables[0].name
r, c = self._loc_to_rc(lit.variables[1])
state[obj_name] = (r, c)
if lit.predicate.name == "carrying":
person_name = lit.variables[1].name
state["carrying"] = person_name
state = tuple(sorted(state.items())) # make hashable
return state
def _loc_to_rc(self, loc_str):
assert loc_str.startswith("f") and loc_str.endswith("f")
r, c = loc_str[1:-1].split('-')
return (int(r), int(c))
def set_state(self, state):
assert isinstance(state, State), "Do not call set_state"
self._state = state
def get_state(self):
assert isinstance(self._state, State), "Do not call get_state"
return self._state
def reset(self):
internal_state, debug_info = super().reset()
return self._internal_to_state(internal_state), debug_info
def step(self, action):
internal_state, reward, done, debug_info = super().step(action)
state = self._internal_to_state(internal_state)
return state, reward, done, debug_info
def get_successor_state(self, state, action):
internal_state = self._state_to_internal(state)
next_internal_state = super().get_successor_state(internal_state, action)
next_state = self._internal_to_state(next_internal_state)
# Sanity checks
assert state == self._internal_to_state(internal_state)
assert next_internal_state == self._state_to_internal(next_state)
return next_state
def render_from_state(self, state):
internal_state = self._state_to_internal(state)
return super().render_from_state(internal_state)
def check_goal(self, state):
internal_state = self._state_to_internal(state)
return super().check_goal(internal_state)
def test_get_all_possible_transitions():
dir_path = os.path.dirname(os.path.realpath(__file__))
domain_file = os.path.join(dir_path, 'pddl',
'test_probabilistic_domain.pddl')
problem_dir = os.path.join(dir_path, 'pddl', 'test_probabilistic_domain')
env = PDDLEnv(domain_file,
problem_dir,
raise_error_on_invalid_action=True,
dynamic_action_space=True)
obs, _ = env.reset()
action = env.action_space.all_ground_literals(obs).pop()
transitions = env.get_all_possible_transitions(action)
transition_list = list(transitions)
assert len(transition_list) == 2
state1, state2 = transition_list[0][0], transition_list[1][0]
type1 = Type('type1')
type2 = Type('type2')
pred1 = Predicate('pred1', 1, [type1])
pred2 = Predicate('pred2', 1, [type2])
pred3 = Predicate('pred3', 3, [type1, type2, type2])
state1, state2 = (state1,
state2) if pred2('c1') in state2.literals else (state2,
state1)
assert state1.literals == frozenset({
pred1('b2'),
pred3('a1', 'c1', 'd1'),
pred3('a2', 'c2', 'd2'),
pred3('b2', 'd1', 'c1')
})
assert state2.literals == frozenset({
pred1('b2'),
pred2('c1'),
pred3('a1', 'c1', 'd1'),
pred3('a2', 'c2', 'd2'),
pred3('b2', 'd1', 'c1')
})
# Now test again with return_probs=True.
transitions = env.get_all_possible_transitions(action, return_probs=True)
transition_list = list(transitions)
assert len(transition_list) == 2
assert abs(transition_list[0][1] -
0.3) < 1e-5 or abs(transition_list[0][1] - 0.7) < 1e-5
assert abs(transition_list[1][1] -
0.3) < 1e-5 or abs(transition_list[1][1] - 0.7) < 1e-5
assert abs(transition_list[0][1] - transition_list[1][1]) > 0.3
state1, state2 = transition_list[0][0][0], transition_list[1][0][0]
type1 = Type('type1')
type2 = Type('type2')
pred1 = Predicate('pred1', 1, [type1])
pred2 = Predicate('pred2', 1, [type2])
pred3 = Predicate('pred3', 3, [type1, type2, type2])
state1, state2 = (state1,
state2) if pred2('c1') in state2.literals else (state2,
state1)
assert state1.literals == frozenset({
pred1('b2'),
pred3('a1', 'c1', 'd1'),
pred3('a2', 'c2', 'd2'),
pred3('b2', 'd1', 'c1')
})
assert state2.literals == frozenset({
pred1('b2'),
pred2('c1'),
pred3('a1', 'c1', 'd1'),
pred3('a2', 'c2', 'd2'),
pred3('b2', 'd1', 'c1')
})
print("Test passed.")
def integration_test():
dir_path = os.path.dirname(os.path.realpath(__file__))
domain_file = os.path.join(dir_path, 'pddl', 'test_domain.pddl')
problem_file = os.path.join(dir_path, 'pddl', 'test_domain', 'test_problem.pddl')
domain = PDDLDomainParser(domain_file)
problem = PDDLProblemParser(problem_file, domain.domain_name, domain.types,
domain.predicates)
## Check domain
type1 = Type('type1')
type2 = Type('type2')
# Action predicates
action_pred = Predicate('actionpred', 1, [type1])
# Predicates
pred1 = Predicate('pred1', 1, [type1])
pred2 = Predicate('pred2', 1, [type2])
pred3 = Predicate('pred3', 1, [type1, type2, type2])
assert set(domain.predicates.values()) == { pred1, pred2, pred3, action_pred }
assert domain.actions == { action_pred.name }
# Operators
assert len(domain.operators) == 1
operator1 = Predicate('action1', 4, [type1, type1, type2, type2])
assert operator1 in domain.operators
operator = domain.operators[operator1]
# Operator parameters
assert len(operator.params) == 4
assert operator.params[0] == type1('?a')
assert operator.params[1] == type1('?b')
assert operator.params[2] == type2('?c')
assert operator.params[3] == type2('?d')
# Operator preconditions (set of Literals)
assert len(operator.preconds.literals) == 4
assert set(operator.preconds.literals) == { action_pred('?b'), pred1('?b'),
pred3('?a', '?c', '?d'), Not(pred2('?c')) }
# Operator effects (set of Literals)
assert len(operator.effects.literals) == 3
assert set(operator.effects.literals) == { Anti(pred2('?d')), pred2('?c'),
pred3('?b', '?d', '?c')}
## Check problem
# Objects
assert set(problem.objects) == {type1('a1'), type1('a2'), type1('b1'),
type1('b2'), type1('b3'), type2('c1'), type2('c2'), type2('d1'),
type2('d2'), type2('d3')}
# Goal
assert isinstance(problem.goal, LiteralConjunction)
assert set(problem.goal.literals) == {pred2('c2'), pred3('b1', 'c1', 'd1')}
# Init
assert problem.initial_state == { pred1('b2'), pred2('c1'),
pred3('a1', 'c1', 'd1'), pred3('a2', 'c2', 'd2'), action_pred('a1'),
action_pred('a2'), action_pred('b1'), action_pred('b2')}
print("Test passed.")
def _create_problem_file(self, raw_problem_fname, use_cache=True):
if (not use_cache) or (raw_problem_fname not in self._problem_files):
problem_fname = os.path.split(raw_problem_fname)[-1]
problem_fname = problem_fname.split('.pddl')[0]
problem_fname += '_{}_with_diffs_{}.pddl'.format(
self.domain_name, random.randint(0, 9999999))
problem_fname = os.path.join('/tmp', problem_fname)
# Parse raw problem
action_names = [
] # purposely empty b/c we WANT action literals in there
problem_parser = PDDLProblemParser(raw_problem_fname,
self.domain_name.lower(),
self._types, self._predicates,
action_names)
# Add action literals (in case they're not already present in the initial state)
# which will be true when the original domain uses operators_as_actions
init_state = State(problem_parser.initial_state,
problem_parser.objects, None)
act_lits = self._action_space.all_ground_literals(init_state,
valid_only=False)
problem_parser.initial_state = frozenset(
act_lits | problem_parser.initial_state)
# Add 'Different' pairs for each pair of objects
Different = Predicate('Different', 2)
init_state = set(problem_parser.initial_state)
for obj1 in problem_parser.objects:
for obj2 in problem_parser.objects:
if obj1 == obj2:
continue
# if obj1.var_type != obj2.var_type:
# continue
diff_lit = Different(obj1, obj2)
init_state.add(diff_lit)
problem_parser.initial_state = frozenset(init_state)
# Also add 'different' pairs for goal if it's existential
if isinstance(problem_parser.goal, Exists):
diffs = []
for var1 in problem_parser.goal.variables:
for var2 in problem_parser.goal.variables:
if var1 == var2:
continue
diffs.append(Different(var1, var2))
problem_parser.goal = Exists(
problem_parser.goal.variables,
type(problem_parser.goal.body)(
problem_parser.goal.body.literals + diffs))
# If no objects, write a dummy one to make FF not crash.
if not problem_parser.objects:
problem_parser.objects.append("DUMMY")
# Write out new temporary problem file
problem_parser.write(problem_fname)
# Add to cache
self._problem_files[raw_problem_fname] = (problem_fname,
problem_parser.objects)
return self._problem_files[raw_problem_fname]
