Pytrich HTN Planner

Overview

Pytrich is a progressive search planner for total-order problems in Hierarchical Task Network (HTN).

To run the planner, ensure the following dependencies are installed and compiled:

After compiling these, move them to pandaBuilds directory before executing the planner.

If you use Ubuntu 22.04 probably the compiled panda files in the project will serve and you won't need to build them again

Pytrich was only tested in Ubuntu if you have any problem to run in a different distro or SO, contact us, or contribute to our project.

Initialize and Update Submodules

git submodule init
cd htn-benchmarks
git submodule update

Run the Planner

python __main__.py htn-benchmarks/Blocksworld-GTOHP/domain.hddl htn-benchmarks/Blocksworld-GTOHP/p01.hddl

Benchmarks: The benchmarks are included as a submodule from htn-benchmarks.
Supported Planning Type: Currently, only total order planning is supported.
Parser/Grounder: A custom parser/grounder is not yet implemented; instead, PANDA is used as a grounding subroutine.
Scripts: The Script directory contains useful scripts for running experiments (not working at the moment).

The script __main__.py supports the following arguments:

domain and problem ([domain_file] [problem_file]): Path to the domain and problem file in HDDL format.
--sas_file <file>: Path to the problem already in sas plus format (don't require including .HDDL domain and problem files).
-H, --heuristic <heuristic_type> (str): Heuristic to use. Choices are TDG, LMCOUNT. Default is TDG.
-s, --search <search_algorithm> (str): Search algorithm to use. Choices are Astar, Blind. Default is Astar.
-hp <args>: Heuristic parameters. Specify parameters for the selected heuristic. Each heuristic has its own parameters, which can be found in the respective class. Example: -H LMCOUNT -hp "name=\"MyHeuristic\", use_bid=True, use_ord=False"
-tor: Enable Total-Order reachability during grounder post-processing (unpublished).
-ms: Monitor time and memory usage during search.
-ml: Monitor time during landmark generation.
-mg: Enable post-processing grounder logging.

Several new components of this planner are currently under research. Key projects include:

AND/OR Landmark Generation: We are developing what we call Bidirectional Landmarks using a new AND/OR encoding for HTN planning.
TO Landmark Generation: Exploration of Total-Order landmark computation to capture ordering constraints.
TO Grounding: Pruning unreachable regions in Decomposition Space through Total-Order analysis.
Landmark Ordering: Investigating landmark ordering to inform search, inspired by LAMA.
IP Heuristics: Enhancing Integer Programming (IP) heuristics with landmarks for solving HTN planning problems.
Novelty: Novelty for HTN planning.

[ ] DEVELOP: Develop a new HTN parser/grounder.
[ ] ENHANCE: grounding techniques and post-processing methods (to-reachability).