Numeric-to-Classical Planning Translation

Bachelor thesis project

This repository implements a compilation from bounded numeric planning tasks to classical planning tasks, allowing numeric PDDL problems to be solved by unmodified classical planners.

The underlying numeric PDDL-to-SAS translation is based on the numeric Fast Downward translator developed for the IPC 2023 Numeric Track, extended here with an additional component (numeric_ir) that recovers a clean, symbolic representation of numeric preconditions and effects from the translator's internal finite-domain encoding. This repository then adds a further translation stage that compiles the resulting numeric task into a purely classical (propositional, finite-domain) SAS+ task, solvable by any standard classical planner such as Fast Downward.

Requirements

• Python 3.10+
• A classical planner capable of reading plain SAS+ input, e.g. Fast Downward (built separately; not included in this repository)

Usage

Batch (all registered instances)

on path src/translate

python3 translate.py

With no arguments, translates every instance registered in ALL_BETA (currently drone1–drone20 and expedition1–expedition20), reading each domain/problem pair from pddl/, and writes one output_classical_<instance>.sas per instance. Useful for local testing of the full benchmark set without submitting a batch job.

Example domain/problem pairs are provided in instances/.

How the pipeline runs the classical search

run_pipeline.py handles the entire process end to end for a single instance: it translates the numeric PDDL task, then automatically pipes the resulting classical SAS+ task into Fast Downward's search component and records the result. You do not need to invoke the search binary yourself — running

python3 run_pipeline.py drone1

performs translation and search in one step, writing output_classical_drone1.sas, sas_plan_drone1, and results/drone1.json.

Internally, this is equivalent to running:

/path/to/fast-downward/builds/release/bin/downward \
    --search "astar(blind())" \
    --internal-plan-file sas_plan_drone1 \
    < output_classical_drone1.sas

Output: run_pipeline.py automatically parses the search binary's stdout and records status (solved/unsolvable/timeout), search time, peak memory, states expanded, and plan length into results/<instance>.json. Running report.py afterward consolidates these across every evaluated instance into results.csv, a summary table, and scaling plots.

Search configuration: astar(blind()) performs optimal A* search using the blind heuristic --- the simplest admissible heuristic, giving A* no guidance beyond the task's own structure.

Notes

Numeric variable bounds must be supplied manually per instance (see ALL_BETA in translate.py), since the numeric SAS+ input format does not itself encode a bound for numeric variables. See the thesis text for a discussion of this design choice and its limitations.