RVRepLan

RVRepLan is a prototype implementation of runtime verification guided replanning. It combines PDDL planning, generated Dejavu monitors, and domain-specific connectors to detect when an execution diverges from the expected action preconditions or effects and to trigger replanning from the current snapshot.

Repository structure

• abstract_system.py: orchestration entry point for planning, monitor synthesis, execution monitoring, issue detection, and replanning.
• translators/: conversion utilities from PDDL action definitions and plans to the monitor properties consumed by Dejavu.
• connectors/: interfaces and benchmark connectors that simulate interaction with the external system.
• mappers/: translation layer between planner actions/perceptions and connector commands/propositions.
• data/: small data model for actions, propositions, and the current execution snapshot.
• benchmarks/: PDDL domains and problems used in the experiments.
• experiments/: experiment outputs and stored plans used for stability analysis.

What to install

• Download and Install DEJAVU (https://github.com/havelund/dejavu) and be sure it works [tested on version 1.0].
• Download and Install Fast-Downward (https://www.fast-downward.org) and be sure it works [tested on version 22.12+]
• Download and Install Python 3 and be sure it works [tested on version 3.10.12].

How to run

To run RVReplan, open a terminal in the installation folder and run:

python3 abstract_system.py DOMAIN PROBLEM CONNECTOR DEJAVU PLANNER

where:

• abstract_system.py: the main script which implements RVReplan algorithms
• DOMAIN: the domain file (.pddl) describing the domain of interest
• PROBLEM: the problem file (.pddl) describing the initial state of the system (i.e., the problem to solve)
• CONNECTOR: the connector used by RVReplan to connect with the real system (the connectors for the three domains used in the experiments are available)
• DEJAVU: the path to the folder where Dejavu has been installed (where dejavu.jar is)
• PLANNER: the path to the planner and any flags it may use

Example of use for the three different domains experimented in the paper:

Logistics:

python3 abstract_system.py ./benchmarks/logistics/domain.pddl ./benchmarks/logistics/problem.pddl connectors.raw_connector_logistics DEJAVU "PLANNER/fast-downward.py --alias lama-first"

python3 abstract_system.py ./benchmarks/logistics/domain.pddl ./benchmarks/logistics/problem.pddl connectors.raw_connector_logistics DEJAVU "PLANNER/fast-downward.py --alias seq-opt-bjolp"

python3 abstract_system.py ./benchmarks/logistics/domain.pddl ./benchmarks/logistics/problem.pddl connectors.raw_connector_logistics DEJAVU "PLANNER/fast-downward.py --alias seq-opt-lmcut"

Depot:

python3 abstract_system.py ./benchmarks/depot/domain.pddl ./benchmarks/depot/problem.pddl connectors.raw_connector_depot DEJAVU "PLANNER/fast-downward.py --alias lama-first"

python3 abstract_system.py ./benchmarks/depot/domain.pddl ./benchmarks/depot/problem.pddl connectors.raw_connector_depot DEJAVU "PLANNER/fast-downward.py --alias seq-opt-bjolp"

python3 abstract_system.py ./benchmarks/depot/domain.pddl ./benchmarks/depot/problem.pddl connectors.raw_connector_depot DEJAVU "PLANNER/fast-downward.py --alias seq-opt-lmcut"

Satellite:

python3 abstract_system.py ./benchmarks/satellite/domain.pddl ./benchmarks/satellite/problem.pddl connectors.raw_connector_satellite DEJAVU "PLANNER/fast-downward.py --alias lama-first"

python3 abstract_system.py ./benchmarks/satellite/domain.pddl ./benchmarks/satellite/problem.pddl connectors.raw_connector_satellite DEJAVU "PLANNER/fast-downward.py --alias seq-opt-bjolp"

python3 abstract_system.py ./benchmarks/satellite/domain.pddl ./benchmarks/satellite/problem.pddl connectors.raw_connector_satellite DEJAVU "PLANNER/fast-downward.py --alias seq-opt-lmcut"

where DEJAVU and PLANNER (replace .py with .sif if using the container version of FD) need to be replaced with their respective paths.

To simulate failures and cause replanning

To ask RVReplan to simulate failures in order to cause replanning it is enough to pass an additional flag:

python3 abstract_system.py DOMAIN PROBLEM CONNECTOR DEJAVU PLANNER --inject_errors N

where the N parameter denotes the kind of failures to introduce:

• 0 no failures (this is the default if no --inject_errors flag is passed)
• 1 one failure on the precondition of an action
• 2 one failure on the postcondition of an action
• 3 both previous failures are added Note that, the addition of a failure may cause more replanning calls.

Other useful flags

In addition to the other parameters and flags, it is also possible to pass:

• --no_monitor_synthesis: this flag causes RVReplan to not synthesise a new monitor, but to use an existing one (i.e., the one available in the folder); this is useful when running multiple times on the same domain, since the monitor would be the same.

To run experiments

To run the experiments on all benchmarks using all three Fast-Downward options (lama-first, bjolp, and lmcut).

sh ./experiments.sh

Once the script is done, you should find in each sub-folder of the experiments folder the CSV containing the experimental results.

Reproducibility notes

The prototype expects external installations of Dejavu and Fast-Downward. The paths in experiments.sh are local placeholders and should be adapted before running the experiments on another machine. Generated runtime artifacts are written under out/ during execution, while replanning traces and CSV metrics are stored in the corresponding experiment folders.