PDDLFuse: A Tool for Generating Diverse Planning Domains

Why it matters

• Planning models are typically trained and evaluated on a limited set of manually created or reconstructed domains.
• Limited domain diversity restricts generalization and encourages overfitting to benchmark distributions (for example, IPC-style domains).
• Domain reconstruction with LLMs often reproduces known domains rather than creating structurally novel ones.
• Reinforcement learning shows domain randomization improves robustness, but planning lacks an equivalent systematic mechanism.
• Robust evaluation of planners and planning foundation models requires scalable generation of diverse and structurally complex domains.

What we did

• Introduced PDDLFuse, a tool that generates new PDDL domains by fusing two existing domains.
• Applied probabilistic modifications to action preconditions and effects (add, remove, and negate predicates) to control structural variation.
• Generated solvable problem instances by executing random action sequences from initial states.
• Evaluated solvability across increasing fusion depths (0-5), showing performance drops for LPG (10/10 to 4/10) and GOOSE (9/10 to 2/10), while the AVI-based model maintains 7/10 at Depth 5.
• Measured structural complexity using the k-WL test, where at Depth 5, 55% of domains require 3-WL and 20% require greater than 3-WL.

As shown in Table 1, increasing fusion depth systematically challenges both traditional planners and foundation models.

How it works

• Domain fusion: merge objects, predicates, and actions from two base domains with systematic renaming to prevent overlap.
• Probabilistic modification: add, remove, and negate predicates in preconditions and effects using configurable probabilities.
• Problem generation: execute random action sequences to derive solvable goal states.
• Complexity analysis: measure structural expressivity using k-WL tests across increasing depth levels.
• Planner evaluation: test FD(FF), LPG, GOOSE, and an AVI-based model using Batch Weighted A* search.

Key contributions

• A configurable tool for generating novel PDDL domains via domain fusion rather than reconstruction.
• Empirical evidence that increasing fusion depth reduces solvability for domain-independent planners (for example, LPG declines from 10/10 at Depth 1 to 4/10 at Depth 5).
• Demonstration that training on PDDLFuse-generated domains improves generalization for an AVI-based model compared to GOOSE at higher depths.
• WL-based expressivity analysis showing that at Depth 5, 55% of domains require 3-WL and 20% require greater than 3-WL, indicating increasing structural complexity.

Figure 1 shows that deeper fused domains increasingly require higher-order WL tests.