Marco Roveri

The goal is to tackle the Multi-Agent Path Finding problem, where multiple agents must coordinate to reach a common goal, using advanced machine learning techniques. In this case, we aim to utilize Hebbian learning to develop a model for each agent that can learn from execution and continuously improve.
Given the complex and sensitive nature of this problem, the network will also integrate logical constraints to ensure the solution's correctness.

The thesis focuses on fine-tuning a large language model for planning tasks. In planning, one has to create a knowledge base composed not only of the predicates describing the environment, but also of the actions that the agents involved in the plan can carry out. This task is quite challenging as one has to consider the different predicates involved in the actions and how they change the state of the plan. Out-of-the-box LLMs provide a good initial guess, but they are still far from providing a knowledge base that is complete and functional on the first iteration. One improvement would be to fine-tune an LLM to improve the results obtained. Moreover, engineering a feedback loop so that the LLM can update the output automatically and fixing the errors may provide even better results.

I have many theis available, get in touch for the complete list.

The scope of this thesis is to design and develop a secure update functionality for IoT devices that will include in the certificate a proof of the correctness of the update w.r.t. some verification condition.

In the IoT security setting, GlobalPlatform plays an important role (https://globalplatform.org/). However, no proper testing infrastructure with coverage guarantees has been developed yet. The aim is to develop such an infrastructure to enable for structured testing of the GlobalPlatform.

The aim is to develop a strong LTLf synthesis tool to automatically synthesize a program that will adhere to the LTLf specification.

Online bin-packing is a problem in which you need to make an immediate decision about the placement of items of various size into fixed capacity bins. The decision can be based on a policy in which each bin that can be chosen is scored based on its remaining capacity and the size of the current item to be placed. The item is then placed into the bin with the highest score. The goal is to maximize the average bins fullness.

The proposed thesis aim to formalize a complex problem as an online bin-packing problem and to solve it with traditional methods and with the usage of metaheuristic techniques. 
 

This thesis, to be carried out together with an SME, aims to build an AI system to support the marketing and accounting offices in preparing quotes/estimates of services/gools.

This thesis aims to link GenAI (e.g. ChatGPT) to control a robotic system like the ARI robot in interacting with humans.

The increasing adoption of robotic applications collaborating with
humans within industrial (e.g. manufacturing, intra-logistics) and
civil (e.g. hospital, pharmaceutical) environments is demanding for
the development of new algorithms, extending state-of-the-art AI
techniques to cope with the new competing requirements: safety for the
interaction with humans; speed to improve productivity.

Artificial Intelligence operational planning, scheduling and actuation
(motion planning) of advanced robotic platforms (AGV, UAV, ...) are
the mean to orchestrate the robotic operations in collaborating with
human to achieve complex business relevant objectives like e.g. reduce
time-span, optimize resource consumption, quickly adapt to
contingencies (e.g. product and/or platform faults, changes in
objectives, changes in human operations). As the capabilities
(e.g. dexterity) of robotic agents grows and their adoption in solving
complex tasks collaborating with humans increases together with an
increase in the complexity of the objectives, the problems of finding
feasible plans to orchestrate the different agents, and of dynamically
adapt and react to run-time contingencies to fulfill the high level
objectives are becoming more and more difficult.

The main tasks for the PhD student are to work on algorithms and
artificial intelligence techniques from the fields of constraint
programming, mathematical optimization, path planning, evolutionary
algorithms and/or reinforcement learning i) to solve such planning,
scheduling, actuation (motion planning) and reactive adaptation
problems; ii) to deepen our understanding of these techniques; iii)
and to show how they can be successfully combined to solve problems of
industrial and civil relevance. Moreover, the PhD student will be
allowed to experiment the developed techniques and algorithms in
realistic-size facilities (e.g. Robotic Labs at the University, and
industrial facilities joint with the University of Trento) equipped
with state-of-the-art robotic and automation platforms.

The PhD student will also experiment his/her results in realistic
industrial environments identified in SMEs located in the internal or
marginalized areas of Italy.