Write your planning task as Python classes, then translate
them to PDDL files. We will use PDDL 1.2 with (:requirements :strips :typing)
, i.e.
the actions only use positive preconditions and deterministic effects, and
we use 'types' like in OOP to represent sets of objects.
The library is written with these considerations:
- Type checking
- To define the domain and problem as Python but yet still look similar to PDDL
- To quickly define the domain and problem using some boilerplate
- To define ground predicates cleanly
- To reduce the use of strings while defining ground predicates
This library allows you to define:
- Domain: requirements (strips and typing), types, predicates, action
- Problem: objects, init, goal
Negation is done using the ~
operator.
- Python 3.6
- python-fire (
pip install fire
)
pip install git+https://github.com/remykarem/py2pddl#egg=py2pddl
We will use the following air cargo problem:
Create an aircargo.py
file by running:
python -m py2pddl.init aircargo.py
and enter the following:
Name: AirCargo
Types (separated by space): cargo airport plane
Predicates (separated by space): plane_at cargo_at in_
Actions (separated by space): load unload fly
In the aircargo.py
source file, the AirCargoDomain
class has been created.
The structure of the class is similar to how a PDDL domain should be defined.
- Name of the domain is the name of the Python class (
AirCargoDomain
). - Types are defined as class variables at the top (
Plane
,Cargo
,Airport
). - Predicates are defined as instance methods decorated with
@predicate
. - Actions are defined as instance methods decorated with
@action
.
Now, complete the class definition such that it looks like this:
from py2pddl import Domain, create_type
from py2pddl import predicate, action
class AirCargoDomain(Domain):
Plane = create_type("Plane")
Cargo = create_type("Cargo")
Airport = create_type("Airport")
@predicate(Cargo, Airport)
def cargo_at(self, c, a):
"""Complete the method signature and specify
the respective types in the decorator"""
@predicate(Plane, Airport)
def plane_at(self, p, a):
"""Complete the method signature and specify
the respective types in the decorator"""
@predicate(Cargo, Plane)
def in_(self, c, p):
"""Complete the method signature and specify
the respective types in the decorator"""
@action(Cargo, Plane, Airport)
def load(self, c, p, a):
precond = [self.cargo_at(c, a), self.plane_at(p, a)]
effect = [~self.cargo_at(c, a), self.in_(c, p)]
return precond, effect
@action(Cargo, Plane, Airport)
def unload(self, c, p, a):
precond = [self.in_(c, p), self.plane_at(p, a)]
effect = [self.cargo_at(c, a), ~self.in_(c, p)]
return precond, effect
@action(Plane, Airport, Airport)
def fly(self, p, orig, dest):
precond = [self.plane_at(p, orig)]
effect = [~self.plane_at(p, orig), self.plane_at(p, dest)]
return precond, effect
Note:
- To create a new type
Car
, simply addCar = create_type("Car")
at the top of the class definition. - The positional arguments of
@predicate
and@action
decorators are the types of the respective arguments. - Methods decorated with
@predicate
should have empty bodies. - Methods decorated with
@action
return a tuple of two lists.
At the bottom part of aircargo.py
, there is another class called AirCargoProblem
.
Again, the structure of the class is similar to how a PDDL problem should be defined.
- Name of the domain is the name of the Python class (
AirCargoProblem
). - Objects are defined as the instance attributes in the
__init__
method. - Initial states are defined as a methods decorated with
@init
. - Goal is defined as an instance methods decorated with
@goal
.
Complete the class definition as follows:
from py2pddl import goal, init
class AirCargoProblem(AirCargoDomain):
def __init__(self):
super().__init__()
self.cargos = AirCargoDomain.Cargo.create_objs([1, 2], prefix="c")
self.planes = AirCargoDomain.Plane.create_objs([1, 2], prefix="p")
self.airports = AirCargoDomain.Airport.create_objs(["sfo", "jfk"])
@init
def init(self):
at = [self.cargo_at(self.cargos[1], self.airports["sfo"]),
self.cargo_at(self.cargos[2], self.airports["jfk"]),
self.plane_at(self.planes[1], self.airports["sfo"]),
self.plane_at(self.planes[2], self.airports["jfk"]),]
return at
@goal
def goal(self):
return [self.cargo_at(self.cargos[1], self.airports["jfk"]),
self.cargo_at(self.cargos[2], self.airports["sfo"])]
Note:
- The Python objects (
cargos
,planes
andairports
) are created using the respective types defined in theAirCargoDomain
. For example,AirCargoDomain.Cargo.create_objs([1, 2], None, "c")
will create a Python dictionary{1: AirCargoDomain.Plane("p1"), 2: AirCargoDomain.Plane("p2")}
. This allows cleaner access to these objects while defining initial state and goal, which usually can get pretty messy. - The PDDL objects defined in the
__init__
are meant to be used across the 2 instance methods. - Any method decorated with
@init
must return a list. - Any method decorated with
@goal
must return a list.
-
Generate the PDDL files from the command line by runnning
python -m py2pddl.parse aircargo.py
-
You can also import the parsing function from the module
from py2pddl import parse parse("aircargo.py")
-
The class itself also contains methods to generate the domain and problem PDDL files separately. These methods were inherited from
Domain
.from aircargo import AirCargoProblem problem = AirCargoProblem() problem.generate_domain_pddl() problem.generate_problem_pddl()
Here is the generated domain.pddl
file.
(define
(domain somedomain)
(:requirements :strips :typing)
(:types
airport
cargo
plane
)
(:predicates
(cargo_at ?c - cargo ?a - airport)
(in_ ?c - cargo ?p - plane)
(plane_at ?p - plane ?a - airport)
)
(:action fly
:parameters (?p - plane ?orig - airport ?dest - airport)
:precondition (plane_at ?p ?orig)
:effect (and (not (plane_at ?p ?orig)) (plane_at ?p ?dest))
)
(:action load
:parameters (?c - cargo ?p - plane ?a - airport)
:precondition (and (cargo_at ?c ?a) (plane_at ?p ?a))
:effect (and (not (cargo_at ?c ?a)) (in_ ?c ?p))
)
(:action unload
:parameters (?c - cargo ?p - plane ?a - airport)
:precondition (and (in_ ?c ?p) (plane_at ?p ?a))
:effect (and (cargo_at ?c ?a) (not (in_ ?c ?p)))
)
)
And here is the generated problem.pddl
file.
(define
(problem someproblem)
(:domain somedomain)
(:objects
sfo jfk - airport
c1 c2 - cargo
p1 p2 - plane
)
(:init (cargo_at c1 sfo) (cargo_at c2 jfk) (plane_at p1 sfo) (plane_at p2 jfk))
(:goal (and (cargo_at c1 jfk) (cargo_at c2 sfo)))
)
Then use your favourite planner like Fast Downward. To output a plan. Here's the plan generated from the above PDDL:
(load c1 p1 sfo)
(fly p1 sfo jfk)
(load c2 p1 jfk)
(unload c1 p1 jfk)
(fly p1 jfk sfo)
(unload c2 p1 sfo)
; cost = 6 (unit cost)
See more examples in the pddl/
folder.
If you want the problem PDDL to be more dynamic if you have
changing inits and goals, you could use dictionaries and
specify in the init
or goal
keyword argument.
p.generate_problem_pddl(
goal={"cargo": "C2"})
Below are several example domains. The respective Python files, PDDL files and sas_plan files (generated using Fast Downward) can be found in the pddl/
folder here.
If you use this software for your work, please cite us as follows:
@article{bin_Karim_py2pddl_2020,
author = {bin Karim, Raimi},
journal = {https://github.com/remykarem/py2pddl},
month = {11},
title = {{py2pddl}},
year = {2020}
}