---
title: Planning Domain Definition Language (PDDL)
url: https://www.emergentmind.com/topics/planning-domain-definition-language-pddl
type: topic
---

# Planning Domain Definition Language (PDDL)

The Planning Domain Definition Language (PDDL) is the predominant formalism used in artificial intelligence to specify symbolic planning domains and problems. Developed originally to standardize input for the International Planning Competitions, PDDL now serves as the foundational modeling language for research and deployment in automated planning and classical, temporal, and hybrid reasoning tasks. PDDL achieves a careful separation between *domain* (defining actions, predicates, types) and *problem instance* (defining concrete objects, initial state, and goals), offering a logic-based, extensible framework with backward compatibility across major planning paradigms [1106.4561][2303.16967][2411.08040].

## 1. Syntax, Structure, and Language Design

PDDL separates the *domain*—the general specification of available actions and object types—from the *problem instance*—a concrete scenario defined by object instantiations, the initial state, and the goal. The canonical PDDL domain contains:

- `:requirements` listing key language features (e.g., `:strips`, `:typing`, `:fluents`, `:durative-actions`).
- `:types` to define a type hierarchy for objects.
- `:predicates` for propositional relations.
- `:functions` for numeric fluents (from PDDL2.1 onward).
- `:action` schemas specifying parameterized instantaneous or durative actions, each with their own preconditions and effects.
- Optional extensions for derived predicates/axioms, processes, events, and context (e.g., for semantic web integration).

A problem file instantiates objects, specifies the `:init` state as a conjunction of ground atoms and numeric values, and sets a conjunctive, disjunctive, or quantified `:goal` condition. Plan metrics are expressed via a `:metric` field, enabling cost, time, or arbitrary numeric objectives [1106.4561][2303.16967][2411.08040][1110.2200].

## 2. Extensions for Temporal, Numeric, and Hybrid Planning

PDDL has evolved via major extensions:

- **PDDL2.1** introduced numeric fluents, arbitrary arithmetic in preconditions/effects, plan metrics, and durative actions. Durative actions can specify temporal conditions (`at start`, `over all`, `at end`) and continuous effects using the special variable `#t` to denote elapsed time. Plan validity requires concurrency and mutual exclusion (mutex) checks [1106.4561].
- **PDDL+** adds *processes* (autonomous, precondition-governed continuous flows) and *events* (uncontrolled, instantaneous state changes triggered by thresholds), mapping the semantics to hybrid automata. Every planning instance can be interpreted as a hybrid automaton over finite locations and continuous variables, yielding a plan validation problem closely related to reachability in control theory [1110.2200][2303.16967].
- **Derived predicates (axioms)** encode intensional relations evaluated as least fixed points. Negations of derived predicates are restricted for decidability: PDDL2.2 allows only negation of basic predicates in axiom bodies, though a stratified-program transformation can eliminate negative derived predicate occurrences [2510.14412].

## 3. Automated Domain and Problem Generation

Recent advances harness LLMs and VLMs to automate generation of PDDL domains, problems, and even abstractions from raw perceptual inputs:

- **SPAR** leverages LLMs to synthesize PDDL domains for UAV tasks from natural language, employing retrieval-augmented chain-of-thought prompting, action-by-action iterative generation, and LLM-friendly syntax validation, attaining high structural and semantic feasibility rates [2509.13691].
- **Image2PDDL** uses vision-language models to parse initial state images and textual or visual goals, then outputs fully specified PDDL problems for domains such as blocksworld and sliding puzzle, demonstrated to achieve 100% syntax validity and very high content fidelity in experiments [2501.17665].
- **PDDLego+** and end-to-end LLM agentic frameworks enable zero-shot translation of textual environmental descriptions into PDDL, iterative refinement, and robust performance in partially observable or ambiguous settings. These pipelines yield interpretable, transferable PDDL representations and maintain strong goal-reaching rates even as task complexity increases [2505.13126][2512.09629].
- **Symbol emergence**: Intrinsically motivated learning agents can autonomously discover symbolic options and abstract them into PDDL fluents and actions, fully automating the mapping from low-level closed-loop controllers to symbolic planning domains [1907.08313].

## 4. Interoperability, Semantics Integration, and Verification

The language’s modular design enables coupling with external reasoning infrastructures:

- **Linked PDDL / PDDLS** introduces a `:context` section for symbol-IRI mapping, allowing seamless integration with semantic web tools (OWL, SPARQL, SHACL). Planners may infer commonsense constraints and conditional facts from ontologies, auto-populating initial states and action preconditions [1912.07834].
- **SysML-PDDL integration**: Profiles bridge system engineering models (SysML v1.6) to automated PDDL generation, enforcing syntactic/semantic constraints via OCL, and supporting "correct-by-construction" planning pipelines in engineering applications [2506.06714].
- **Programmatic and sound LLM planning**: By synthesizing policies and heuristics as executable programs over PDDL state spaces, planners can guarantee soundness by construction. Such approaches achieve competitive or superior coverage to classic planners and exhibit some robustness to symbol "gensymming," challenging the view that LLMs merely memorize from English [2508.18507].
- **Universal domains**: Construction of a single, lifted PDDL domain—parameterized only by static predicates—can encode any propositional planning instance, revealing that generalised planning is as hard as instance-by-instance planning for the universal domain. No polynomial-size truly lifted universal exists under arity bounds [2411.08040].

## 5. Knowledge Engineering Tools and Automated Validation

PDDL’s widespread adoption is supported by specialized tooling:

- **myPDDL** supplies modular project scaffolding, snippets, context-aware syntax highlighting, type-hierarchy visualization, and integration with host languages for preprocessing (e.g., Clojure-driven distance calculation). Controlled studies confirm these tools reduce error rates (by 36%) and accelerate comprehension and debug time (by 48%) for both novices and expert engineers [2008.11069].
- **Benchmarks and evaluation frameworks**: Datasets like Planetarium and LLM evaluation studies provide rigorous semantic-equivalence algorithms for scoring text-to-PDDL generation. Despite high syntax and solvability rates (e.g., 96%+ parseability), semantic correctness in generated PDDL remains a bottleneck for LLMs, with top models achieving only 24.8–35.1% strict correctness on realistic tasks [2407.03321][2502.20175].

## 6. Advanced Modeling: Streams, Fusion, and Diversity

- **PDDLStream** generalizes the language with `:streams`: declarative interfaces for integrating blackbox samplers, enabling task and motion planning in high-dimensional, continuous, or sampled spaces. This allows joint synthesis of symbolic and geometric/kinematic plans using optimistic, adaptive interleaving of symbolic search and sampling [1802.08705].
- **Domain randomization and fusion**: Tools such as PDDLFuse systematically generate diverse, random planning domains by fusing predicates and actions from existing domains with controlled parameterization, enabling scalable tests of planner generalizability and robustness. Empirical results show that increasing fusion depth and the extremity of randomization parameters dramatically affect planner solvability and path cost, illustrating PDDL’s use in driving both research benchmarking and meta-planning methodologies [2411.19886].

## 7. Theoretical Foundations and Computational Complexity

PDDL-based planning is subsumed by the classical complexity of propositional STRIPS planning—-plan existence is PSPACE-complete. The introduction of numeric, temporal, or hybrid constructs increases expressive power but may result in undecidable plan existence in general. PDDL+ domains map to hybrid automata, inheriting their complexity/decidability frontier: reachability is undecidable in the general case, but certain fragments (priced timed automata, initialized singular automata) remain decidable [1110.2200][2411.08040].

## References

- [1106.4561] PDDL2.1: An Extension to PDDL for Expressing Temporal Planning Domains
- [1110.2200] Modelling Mixed Discrete-Continuous Domains for Planning
- [1907.08313] Learning High-Level Planning Symbols from Intrinsically Motivated Experience
- [1912.07834] Design and Implementation of Linked Planning Domain Definition Language
- [2008.11069] MyPDDL: Tools for efficiently creating PDDL domains and problems
- [2303.16967] Heuristic Search For Physics-Based Problems: Angry Birds in PDDL+
- [2407.03321] Planetarium: A Rigorous Benchmark for Translating Text to Structured Planning Languages
- [2411.08040] The Universal PDDL Domain
- [2411.19886] PDDLFuse: A Tool for Generating Diverse Planning Domains
- [2501.17665] Planning with Vision-Language Models and a Use Case in Robot-Assisted Teaching
- [2502.20175] An Extensive Evaluation of PDDL Capabilities in off-the-shelf LLMs
- [2505.13126] Zero-Shot Iterative Formalization and Planning in Partially Observable Environments
- [2506.06714] Integrating AI Planning Semantics into SysML System Models for Automated PDDL File Generation
- [2508.18507] Language Models For Generalised PDDL Planning: Synthesising Sound and Programmatic Policies
- [2509.13691] SPAR: Scalable LLM-based PDDL Domain Generation for Aerial Robotics
- [2510.14412] Eliminating Negative Occurrences of Derived Predicates from PDDL Axioms
- [2512.09629] An End-to-end Planning Framework with Agentic LLMs and PDDL
- [1802.08705] PDDLStream: Integrating Symbolic Planners and Blackbox Samplers via Optimistic Adaptive Planning

Source: https://www.emergentmind.com/topics/planning-domain-definition-language-pddl