Graph Learning for Planning: The Story Thus Far and Open Challenges

TL;DR

This paper surveys graph-based Learning for Planning (L4P), framing the problem as learning domain knowledge from small planning tasks to scale to larger ones. It introduces the GOOSE framework and argues that classical ML methods, particularly Weisfeiler-Lehman (WL) feature-based approaches, often outperform deep learning models in symbolic planning, while ranking-based formulations yield stronger, more reliable guidance for search. The authors provide a taxonomy of graph representations (grounded, IR, PR) and compare expressivity and practical performance, highlighting that expressivity, generalisation, and data collection are central open challenges. The work underscores that simple, fast, data-efficient graph methods can rival strong planners in many settings and sets a research agenda focused on expressivity, generalisation, optimisation, data collection, and fair benchmarking. Its findings have practical implications for building scalable, domain-agnostic planning systems that leverage relational structure efficiently.

Abstract

Graph learning is naturally well suited for use in planning due to its ability to exploit relational structures exhibited in planning domains and to take as input planning instances with arbitrary number of objects. In this paper, we study the usage of graph learning for planning thus far by studying the theoretical and empirical effects on learning and planning performance of (1) graph representations of planning tasks, (2) graph learning architectures, and (3) optimisation formulations for learning. Our studies accumulate in the GOOSE framework which learns domain knowledge from small planning tasks in order to scale up to much larger planning tasks. In this paper, we also highlight and propose the 5 open challenges in the general Learning for Planning field that we believe need to be addressed for advancing the state-of-the-art.

Figures (8)

• Figure 1: Typical graph learning for planning pipeline. (1) Planning tasks are represented as graphs. (2) Graph learning architectures are used to embed graphs. (3) Learning is performed via some optimisation criteria suited for making predictions in planning.
• Figure 2: Visualisations of generalisation setups for planning and RL. The axis represents the number of objects in a planning task. Standalone planning and RL commonly follow the task-specific setup (left), with some RL approaches generalising across similar-sized tasks (middle). L4P approaches primarily focuses on generalisation across arbitrary numbers of objects (right).
• Figure 3: The general Learning for Planning (L4P) setup.
• Figure 4: Expressivity hierarchy of graph representations of planning tasks. Colours of edges denote the key takeaways classified in Section \ref{['ssec:hierarchy']}.
• Figure 5: Time to train after data preprocessing in log scale (left) and number of learnable parameters (right) of GNN and WL models on various planning domains.