Title:	Advancing Neuro-symbolic Approaches for Automated Planning
DNr:	Berzelius-2026-49
Project Type:	LiU Berzelius
Principal Investigator:	Jendrik Seipp <jendrik.seipp@liu.se>
Affiliation:	Linköpings universitet
Duration:	2026-02-12 – 2026-09-01
Classification:	10210
Keywords:

Abstract

Automated planning is the task of finding sequences of actions that achieve specified goals, a fundamental problem in AI with applications ranging from robotics to logistics and energy optimization. A common example is the game “Tower of Hanoi” where the goal is to move an entire stack of disks, arranged in decreasing size, from one peg to another. Despite recent advances in both symbolic AI and deep learning, automated planning remains computationally challenging, particularly for NP-hard problems that arise in real-world applications, as the international planning competition (IPC) shows (https://ipc2023-learning.github.io/). While traditional symbolic planners provide guarantees and interpretability, they often struggle with scalability. Conversely, learned approaches can generalize across problem instances but lack the robustness and theoretical guarantees of symbolic methods. Unlike traditional reinforcement learning, automated planning focuses on a class of problems called a domain. For example, in the Tower of Hanoi domain, there exists an algorithm that solves all instances independently of the number of disks and pegs. Recent trends in reasoning, especially for LLM-based methods, demonstrate an increased interest in such tasks (https://arxiv.org/abs/2501.12948). However, the vast majority of such approaches do not target out-of-distribution generalization, which we aim to address by transferring knowledge from easy to hard problems within the same domain while focusing on agnostic approaches that only require domain and problem descriptions in a standard intermediate presentation language called PDDL. This project addresses the critical research gap of developing neuro-symbolic systems that combine the strengths of both paradigms—leveraging neural networks for efficient generalization while maintaining the logical rigor and interpretability of symbolic reasoning. Understanding how to design such hybrid architectures is essential for advancing AI systems that are both powerful and trustworthy. Research Objectives: - Develop hybrid systems combining large language models with symbolic planners to leverage both semantic understanding and logical reasoning. - Analyze the expressiveness and generalization capabilities of neural architectures (transformers, graph neural networks, etc.) for planning tasks. - Apply logical formalisms to interpret and verify learned representations, ensuring alignment with symbolic planning principles. - Design robust neural architectures that inherently capture algorithmic structures and provide generalization guarantees. - Investigate energy-based probabilistic models for handling out-of-distribution scenarios in automated planning. - Characterize the computational complexity boundaries and limitations of neuro-symbolic methods on NP-hard planning problems. - Train generative models to produce candidate plans, abstractions, or intermediate symbolic representations that can be refined and verified by classical planners. - Train models that allow the extraction of human-interpretable algorithms using deep symbolic regression.