The following projects can be either bachelor (triennale) or master (magistrale) thesis or even PhD opportunities. Some is more specific some is more open depending on the level of involvement of the student. All projects are at the frontier of automated planning and artificial intelligence in general. Students willing to cooperate have the chance to engage with an international research panorama for a better understanding of the science behind the computers and how this can be used to build intelligent systems. Our group is very active scientifically worldwide on all problems related to automated planning, scheduling and Artificial Intelligence in general.
Non exhaustive list of projects:
Title: Powerful Heuristics for State Space Search Planners
Abstract: Development of innovative relaxation-based heuristics in the context of hybrid planning. These heuristics can pave the way of a more prominent use of automated planning in real-world application, from robotics, urban traffic control, games and personalized medication. Attention is going to be paid to those numeric planning problems exhibiting state-dependent numeric dependencies between variables.
Supervisors: Enrico Scala
Title: Declarative Simulators and Controllers through PDDL+
Abstract: Development of an automatic translator from Simulink/Stateflow models (part of the Matlab environment) to PDDL+ planning. The translator aims to bridge the gap between industry prototyping and automated planning by converting dynamic system models into Planning Domain Definition Language (PDDL+) representations. This enables the application of advanced planning techniques to simulate and analyze model behavior. The project will involve identifying a fragment of relevant Simulink/Stateflow operators, defining a translation algorithm, implementing the translator, and validating its effectiveness on a range of models. The work by this thesis holds the promise to enable the automated verification and analysis of complex dynamical systems through the use of automated planning systems.
Supervisors: Diego Aineto, Ivan Serina, Enrico Scala
Title: Action Model Learning for Planning
Abstract: Design and theoretical study of an approach that learns automatically models of the environment expressed using action centered formalisms based on first-order-logic. This work aims at deepening our understanding on what can be learnt and with which guarantees. The action model learning takes as an input some flow of demonstrations that collect the experience of the agent. The ouput is a model that can be used by any off-the-shelf planner. Emphasis will be given to expressive models of the world
Supervisors: Diego Aineto, Enrico Scala
Title: Model-Free Planning
Abstract: Design of an algorithm that solves planning problem with no model description given as an input. The idea is to navigate the search space of the planning problem using a simulator and devise heuristics on the basis of the information collected during search. Heuristics can be either syntactic or the result of an incumbent planning model that is learnt whilst palnning.
Supervisors: Diego Aineto, Enrico Scala
Title: Hybrid Planning with Trajectory Constraints
Abstract: In many domains, agents have to deal with trajectory constraints all along the way, for instance to enforce that the battery is up for the entire course of the activities, or perhaps that the connection link is open while some operation is carried on. To take into account this problem, one can use metric temporal constraints. In this thesis we expect to advance the state of the art in all the technologies fo the automated synthesis of plans that take all trajectory constraints into account too. We are aiming to look at signal temporal logic formulae and PDDL+ representation languages, but other alternatives can be considered.
Supervisors: Diego Aineto, Enrico Scala
Title: Multi Objective Planning
Abstract: In real-world planning applications, defining a single objective function is often challenging. For example, the fastest route to a destination may be significantly more expensive than an alternative route that is only slightly slower but substantially cheaper. To systematically address problems involving multiple competing objectives, one can leverage multi-objective optimisation techniques. These approaches enable planning agents to identify diverse solutions, each representing a different trade-off among relevant objectives.
This project will investigate the integration of multi-objective optimisation within automated planning. Depending on the student’s interests, potential research directions include the design of multi-objective heuristics, the development of multi-objective search algorithms, and the visualisation and analysis of trade-offs among alternative plans.
Supervisors: Enrico Scala
Title: Incremental Automated Planning
Abstract: This project focuses on the study and development of incremental search strategies that can effectively reuse information from previous search episodes. Such capabilities are particularly valuable when improving an existing solution, adapting to changes in the environment, or recovering from unexpected failures.
A central objective of the project will be the investigation of the Anytime D* framework and its application to automated planning problems. The work may involve both theoretical analysis and empirical evaluation, with the goal of understanding how incremental search techniques can improve planning efficiency and responsiveness in dynamic settings.
Supervisors: Enrico Scala
Title: LLM-Assisted Model Generation
Abstract: Large Language Models (LLMs) have recently demonstrated remarkable capabilities in understanding and generating structured representations from natural language descriptions. This project aims to investigate the ability of LLMs to automatically generate Planning Domain Definition Language (PDDL) models from textual specifications.
The generated models will be evaluated in terms of their correctness and usefulness for plan-based agents operating in a variety of tasks. The study will compare different language models and assess their performance against manually constructed ground-truth models as well as alternative automated model-generation approaches. The project may also explore methods for improving model quality through prompting strategies, feedback loops, or hybrid human-in-the-loop approaches.
Supervisors: Enrico Scala