This paper introduces LeapNP (Learning and Planning Framework for Numeric Problems), a lightweight, Python-native framework engineered to support both classical and numeric planning tasks. Designed with a fully modular interface, it specifically aims to facilitate the seamless integration of deep learning methodologies. The design philosophy of LeapNP stems from the observation that traditional planners, while highly efficient, lack the necessary flexibility for experimental research, particularly at the intersection of learning and planning. Most state-of-the-art engines are built as highly optimized, rigid executables that are resistant to internal modification. LeapNP disrupts this paradigm by offering a framework where the entire planning stack is accessible and mutable. Users can seamlessly plug in custom implementations for grounding, define novel state representations, or design bespoke search strategies, thereby enabling a level of integration with learning models that is currently impractical with standard tools. By significantly lowering the engineering barrier, our planner fosters rapid experimentation and accelerates research in neuro-symbolic planning. We also present a comprehensive suite of search algorithms, designed to evaluate different properties of learned heuristics. These include two algorithms designed to exploit batching to maximize inference throughput, and a greedy algorithm meant to test the intrinsic robustness of the learned models, running them as general policies.

In this paper, we investigate the application of heuristics based on Graph Neural Networks (GNNs) to lifted numeric planning problems, an area that has been relatively unexplored. Building upon the GNN approach for learning general policies proposed by Staahlberg et al., we extend the architecture to make it sensitive to the numeric components inherent in the planning problems we address. We achieve this by observing that, although the state space of a numeric planning problem is infinite, the finite subgoal structure of the problem can be incorporated into the architecture, allowing for the construction of only a finite number of nodes. Instead of learning general policies, we train our models to function as a heuristic within a best-first search algorithm. We explore various configurations of this architecture and demonstrate that the resulting heuristics are highly informative and, in certain domains, offer a better trade-off between guidance and computational cost compared to other inductive and deductive heuristics.