STRIPS planner with PDDL-like problem specification in Python
by Gary Doran ([email protected])
PyDDL can be installed using pip:
pip install -e git+https://github.com/garydoranjr/pyddl.git#egg=pyddl
or by running the setup script:
python setup.py install
This is a STRIPS-like planner that I wrote for an AI class assignment. A planning domain and problem is specified in Python, in a manner that is intended to be reminiscent of PDDL. However, only a small subset of PDDL-like constructs are implemented. Planning is performed using A* search, but not too much trouble has been taken to optimize the search efficiency. Therefore, PyDDL is best suited for smaller problems.
A planning domain is specified as a Domain object, and takes a list of
Action objects. Unlike PDDL, predicates do not need to be explicitly defined.
An Action takes a list of arguments similar to that for PDDL. For example, see
the eight-puzzle example script containing the following action:
Action(
'move-up',
parameters=(
('tile', 't'),
('position', 'px'),
('position', 'py'),
('position', 'by'),
),
preconditions=(
('dec', 'by', 'py'),
('blank', 'px', 'by'),
('at', 't', 'px', 'py'),
),
effects=(
neg(('blank', 'px', 'by')),
neg(('at', 't', 'px', 'py')),
('blank', 'px', 'py'),
('at', 't', 'px', 'by'),
),
),
One difference from PDDL here is that parameters are typed using (type, name)
tuples rather than specifying types with predicates.
Similarly, a planning problem is specified with a Problem object, which takes
a planning domain, a dictionary of typed objects, a list of initial predicates,
and a list of goal predicates.
Once a problem has been constructed, it can be passed to the planner function
with optional arguments (e.g. a search heuristic) to generate a plan (if one
exists). The resulting plan is a sequence of Action objects to execute, or
None if no plan exists.
PyDDL supports some basic numeric functions, comparisons, and operations. See the "missionaries and cannibals" problem for example usage.