Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PDDLFuse: A Tool for Generating Diverse Planning Domains

Vedant Khandelwal, Amit Sheth, Forest Agostinelli

TL;DR

PDDLFuse addresses the scarcity and limited diversity of planning domains by introducing DomGenX, a parametric tool that generates novel PDDL domains through controlled domain randomization-inspired methods. It combines two base domains, resolves naming overlaps, and injects stochastic variations in preconditions, effects, and negations, producing solvable problem instances with adjustable difficulty. The study demonstrates that domain diversity affects solver performance, showing variable solvability across depths for reference planners and validating the generated domains with a dedicated validator. This approach broadens the testing ground for planning algorithms, enabling stronger inductive biases and better generalization to unseen domains, with practical implications for developing more robust planners. The work also outlines a rigorous experimental framework, including time-bounded evaluation, parameter sweeps, and heat-map analyses, to systematically assess domain-generator effectiveness and planner robustness.

Abstract

Various real-world challenges require planning algorithms that can adapt to a broad range of domains. Traditionally, the creation of planning domains has relied heavily on human implementation, which limits the scale and diversity of available domains. While recent advancements have leveraged generative AI technologies such as large language models (LLMs) for domain creation, these efforts have predominantly focused on translating existing domains from natural language descriptions rather than generating novel ones. In contrast, the concept of domain randomization, which has been highly effective in reinforcement learning, enhances performance and generalizability by training on a diverse array of randomized new domains. Inspired by this success, our tool, PDDLFuse, aims to bridge this gap in Planning Domain Definition Language (PDDL). PDDLFuse is designed to generate new, diverse planning domains that can be used to validate new planners or test foundational planning models. We have developed methods to adjust the domain generators parameters to modulate the difficulty of the domains it generates. This adaptability is crucial as existing domain-independent planners often struggle with more complex problems. Initial tests indicate that PDDLFuse efficiently creates intricate and varied domains, representing a significant advancement over traditional domain generation methods and making a contribution towards planning research.

PDDLFuse: A Tool for Generating Diverse Planning Domains

TL;DR

PDDLFuse addresses the scarcity and limited diversity of planning domains by introducing DomGenX, a parametric tool that generates novel PDDL domains through controlled domain randomization-inspired methods. It combines two base domains, resolves naming overlaps, and injects stochastic variations in preconditions, effects, and negations, producing solvable problem instances with adjustable difficulty. The study demonstrates that domain diversity affects solver performance, showing variable solvability across depths for reference planners and validating the generated domains with a dedicated validator. This approach broadens the testing ground for planning algorithms, enabling stronger inductive biases and better generalization to unseen domains, with practical implications for developing more robust planners. The work also outlines a rigorous experimental framework, including time-bounded evaluation, parameter sweeps, and heat-map analyses, to systematically assess domain-generator effectiveness and planner robustness.

Abstract

Various real-world challenges require planning algorithms that can adapt to a broad range of domains. Traditionally, the creation of planning domains has relied heavily on human implementation, which limits the scale and diversity of available domains. While recent advancements have leveraged generative AI technologies such as large language models (LLMs) for domain creation, these efforts have predominantly focused on translating existing domains from natural language descriptions rather than generating novel ones. In contrast, the concept of domain randomization, which has been highly effective in reinforcement learning, enhances performance and generalizability by training on a diverse array of randomized new domains. Inspired by this success, our tool, PDDLFuse, aims to bridge this gap in Planning Domain Definition Language (PDDL). PDDLFuse is designed to generate new, diverse planning domains that can be used to validate new planners or test foundational planning models. We have developed methods to adjust the domain generators parameters to modulate the difficulty of the domains it generates. This adaptability is crucial as existing domain-independent planners often struggle with more complex problems. Initial tests indicate that PDDLFuse efficiently creates intricate and varied domains, representing a significant advancement over traditional domain generation methods and making a contribution towards planning research.

Paper Structure

This paper contains 39 sections, 4 figures, 217 tables, 4 algorithms.

Table of Contents

  1. Introduction
  2. Background and Related Works
  3. Planning Domains and Problems
  4. Domain Reconstruction
  5. Generalization in Planning
  6. Domain Randomization and Generalization in Reinforcement Learning
  7. Methods
  8. Domain Generation
  9. Problem File Generation
  10. Detailed Algorithms in Appendices
  11. Results
  12. Planners and Heuristics Evaluation
  13. Fast Forward (FF) Heuristic
  14. Experimental Setup and Parameters
  15. Solvability of Planners Across Depth Levels
  16. ...and 24 more sections

Figures (4)

  • Figure 1: 15 Objects -- Fast Downward Planner
  • Figure 2: 15 Objects -- LPG Planner
  • Figure 3: Solvability heat maps for domains with 15 objects, evaluated using the Fast Downward planner with lmcut heuristic across varied parameter configurations.
  • Figure 4: Solvability heat maps for domains with 15 objects, evaluated using the LPG Planner across varied parameter configurations.