The paper focuses on utilizing Large Language Models (LLMs) for the automated generation of PDDL (Planning Domain Definition Language) domains and problems from natural language descriptions.
The key steps in the proposed pipeline are:
Generating a textual goal plan: The initial user input is used by the LLM to generate a richer textual description of the problem, including the initial state, goal state, and a sequence of actions to reach the goal.
Generating domain markup: The textual description is then used to generate a JSON-based markup, which is syntactically similar but different from PDDL. This step aims to reduce syntax-specific errors that can occur when directly generating PDDL.
Initial consistency checks: The generated JSON markup undergoes a series of consistency checks to identify issues such as wrong parameter types, undeclared parameters, predicate mismatches, etc. The errors found are then provided back to the LLM for correction.
Error correction loop: The LLM is prompted with the identified errors and suggestions for resolving them. This back-prompting process continues until the domain and problem are free of the detected inconsistencies.
Reachability analysis: Finally, the corrected PDDL domain and problem are analyzed for reachability of the goal state. This step identifies issues such as actions that are not involved in the plan or predicates that are missing from the initial state.
The authors evaluate the proposed pipeline on several classical and custom planning domains, including logistics, gripper, tyreworld, household, and pizza cooking. The results show that the framework can generate executable PDDL domains and problems for simpler scenarios, but more complex domains still present challenges in terms of consistency and reachability.
The key contribution of this work is the integration of consistency checks and reachability analysis techniques into the LLM-based generation process, which helps reduce the amount of errors in the generated domains and improves their usability for planning.
To Another Language
from source content
arxiv.org
Głębsze pytania