Tyr: Generalized Planning in C++20 and Python

Tyr is designed to address several challenges in modern planning systems:

1. Unified grounded and lifted planning within a type-safe API.

2. Rapid prototyping through Python bindings with type hints, backed by a high-performance C++ core.

3. Support for expressive numeric planning formalisms across both grounded and lifted reasoning paradigms (see Supported PDDL Features).

4. Integration of learning and reasoning by supporting collections of planning tasks over a shared planning domain.

Technical Overview

• PDDL frontend: Tyr uses Loki to parse, normalize, and translate PDDL input. The parser is implemented with Boost and provides informative error messages for syntactically invalid input. The normalization pipeline largely follows the approach described in Section 4 of Concise finite-domain representations for PDDL planning tasks.

• Datalog engine: Tyr implements a parallel semi-naive Datalog engine for lifted successor generation, axiom evaluation, relaxed planning graph heuristics, and task grounding. Its execution model is synchronous and supports both rule-level and grounding-level parallelism.

• Ground planning: For grounded tasks, Tyr uses data structures inspired by The Fast Downward Planning System to efficiently identify applicable actions in a given state. Grounding often yields substantial performance improvements, although it is not always feasible for large tasks.

• State representation: Tyr statically analyzes domain and problem files and partitions predicates, functions, and related structures into strongly typed categories such as static, fluent, and derived atoms. This design prevents accidental mixing of conceptually different entities. To represent sequences compactly, Tyr uses tree databases of perfectly balanced binary trees, allowing common subsequences to be shared through shared subtrees. As a special case, Tyr synthesizes finite-domain variables for fluent atoms in grounded planning, largely following the method described in Section 5 of Concise finite-domain representations for PDDL planning tasks, enabling more compact storage when grounding is feasible.

• Memory model: Tyr stores generated data in hierarchically structured, geometrically growing buffers. For variable-sized objects, it uses Cista for serialization and zero-copy deserialization. This design allows derived buffers to inherit data from parent buffers without duplication. For example, multiple tasks can share a domain, and multiple workers can share task data.

Getting Started

The library consists of a formalism and a planning component. The formalism component is responsible for representing PDDL entities. The planning component provides functionality for implementing search algorithms, as well as off-the-shelf implementations of eager A*, lazy GBFS, and heuristics such as blind, max, add, and FF. Below is a minimal overview of the Python and C++ APIs for implementing custom search algorithms.

Python Interface

Pytyr is available at PyPI and can be installed with pip install pytyr.

Detailed examples are available in the python/examples directory:

• structures.py – Parse and traverse all planning formalism structures.
• builder.py – Create new planning formalism structures.
• invariants.py – Synthesize invariants, access candidate variable bindings, and match atoms through unification.
• astar_eager.py – Use and customize off-the-shelf search algorithms.
• gbfs_lazy.py – Implement a custom search algorithm from scratch.

The Python interface for implementing search algorithms is:

# Recommended namespace aliases
from pytyr.common import ExecutionContext
import pytyr.formalism.planning as tfp
import pytyr.planning.lifted as tpl  # pytyr.planning.ground also exists

# Parse and translate a task over a domain.
parser = tfp.Parser("domain.pddl")
# Instantiate a lifted task.
task = tpl.Task(parser.parse_task("problem.pddl"))

# Instantiate a single-threaded execution environment.
execution_context = ExecutionContext(1)

# Instantiate the planning objects. Factories assign unique context indices so
# state views from different state repositories hash and compare correctly.
axiom_evaluator_factory = tpl.AxiomEvaluatorFactory()
state_repository_factory = tpl.StateRepositoryFactory()
successor_generator_factory = tpl.SuccessorGeneratorFactory()
axiom_evaluator = axiom_evaluator_factory.create(task, execution_context)
state_repository = state_repository_factory.create(task, axiom_evaluator)
successor_generator = successor_generator_factory.create(task, execution_context, state_repository)

# Get the initial node (state + metric value)
initial_node = successor_generator.get_initial_node()

# Get the labeled successor nodes (sequence of ground action + node)
labeled_successor_nodes = successor_generator.get_labeled_successor_nodes(initial_node)

C++ Interface

The C++ interface for implementing search algorithms is:

#include <tyr/tyr.hpp>

// Recommended namespace aliases.
namespace tfp = tyr::formalism::planning;
namespace tp = tyr::planning;

// Parse and translate a task over a domain.
auto parser = tfp::Parser("domain.pddl");
// Instantiate a lifted task.
auto task = tp::Task<tp::LiftedTag>::create(parser.parse_task("problem.pddl"));

// Instantiate a single-threaded execution environment
auto execution_context = tyr::ExecutionContext::create(1);

// Instantiate the planning objects. Factories assign unique context indices so
// state views from different state repositories hash and compare correctly.
auto axiom_evaluator_factory = tp::AxiomEvaluatorFactory<tp::LiftedTag>();
auto state_repository_factory = tp::StateRepositoryFactory<tp::LiftedTag>();
auto successor_generator_factory = tp::SuccessorGeneratorFactory<tp::LiftedTag>();

auto axiom_evaluator = axiom_evaluator_factory.create(task, execution_context);
auto state_repository = state_repository_factory.create(task, axiom_evaluator);
auto successor_generator = successor_generator_factory.create(task, execution_context, state_repository);

// Get the initial node (state + metric value).
auto initial_node = successor_generator->get_initial_node();

// Get the labeled successor nodes (sequence of ground action + node).
auto labeled_successor_nodes = successor_generator->get_labeled_successor_nodes(initial_node);

Dependencies

Tyr consumes native dependencies from Python packages:

• pyyggdrasil >= 0.0.9 for shared third-party native dependencies.
• pypddl >= 1.0.6 for Loki's PDDL parser library, headers, and CMake package.

The shared workspace layout and general Python/CMake integration pattern are documented in the Planning and Learning build instructions.

Build C++

Install Tyr's native dependency providers into the active Python environment, then configure CMake with their native prefixes:

python -m pip install 'pyyggdrasil>=0.0.9' 'pypddl>=1.0.6'

cmake -S . -B build \
  -DPython_EXECUTABLE="$(python -c 'import sys; print(sys.executable)')" \
  -DPython3_EXECUTABLE="$(python -c 'import sys; print(sys.executable)')" \
  -DCMAKE_PREFIX_PATH="$(python -c 'import os, pyyggdrasil, pypddl; print(os.pathsep.join(map(str, [pyyggdrasil.native_prefix(), pypddl.native_prefix()])))')"

cmake --build build -j4

CMake options:

Option	Default	Description
TYR_BUILD_TESTS	OFF	Build Tyr tests.
TYR_BUILD_EXECUTABLES	OFF	Build Tyr executables.
TYR_BUILD_PROFILING	OFF	Build Tyr profiling targets.
TYR_BUILD_PYTYR	OFF	Build pytyr Python bindings.
TYR_ENABLE_FMT_FORMATTERS	ON	Enable Tyr's public fmt::formatter specializations.
TYR_HEADER_INSTANTIATION	OFF	Enable stronger inlining at higher compile-time cost.
TYR_ENABLE_INNER_PARALLELISM	OFF	Enable inner rule parallelism.
TYR_USE_LLD	ON	Use LLVM lld when available.
TYR_ENABLE_LTO	ON	Enable link-time optimization for optimized builds.
TYR_STATE_STORAGE_POLICY	Tree	State storage backend; accepted values are Tree and Hashset.

Install Tyr from a configured build directory with:

cmake --install build --prefix=<path/to/installation-directory>

More detailed Tyr-specific build instructions are available in docs/BUILD.md.

Build Python

python -m pip install .[test]
pytest python/tests

CMake Integration

The Python package pytyr installs Tyr's native headers, shared library, and CMake package config under pytyr.native_prefix(). Downstream CMake projects should include the native prefixes of pytyr and its native package dependencies in CMAKE_PREFIX_PATH:

cmake -S . -B build \
  -DCMAKE_PREFIX_PATH="$(python -c 'import os, pyyggdrasil, pypddl, pytyr; print(os.pathsep.join(map(str, [pyyggdrasil.native_prefix(), pypddl.native_prefix(), pytyr.native_prefix()])))')"

Tyr exports the tyr::core aggregate target.