This Master's thesis delves into the innovative intersection of generative artificial intelligence, virtual reality (VR), and interactive entertainment. The primary goal is to develop a Passport Control Simulator within the Unity engine, inspired by VR experiences like Nothing to Declare.

The core innovation is the creation of enhanced "virtual humans" (VHs) powered by Large Language Models (LLMs). These VHs will engage in dynamic, unscripted dialogue, moving beyond traditional game development by integrating a full conversational pipeline, including Speech-to-Text (STT) for user input and Text-to-Speech (TTS) for agent responses.

Key Tasks & Research Objectives

The student will first conduct a focused literature review on the state-of-the-art in conversational AI, specifically LLM-driven agents and prompt engineering for personality portrayal. This review will include also state-of-the-art methods to measure user engagement in interactive XR applications.

The practical component involves developing the core VR simulator in Unity, including all 3D environment and interaction mechanics. The central task is to implement the conversational and affective engine. This will require integrating an LLM (potentially leveraging frameworks like LangChain) to create a "role-play architecture," enabling VHs to believably portray distinct emotions and personalities (e.g., a nervous traveler vs. a confident one).

The research aspect will be evaluated through a user study designed by the student. This study will aim to answer:

• Are LLM-based architectures able to believably portray different personalities?
• Can users discern these personalities?
• How is user engagement affected by this advanced, natural-language interaction compared to simple interaction techniques?

Expected Outcomes

• A functional VR prototype in Unity that demonstrates a closed-loop, real-time conversational system with LLM-driven virtual humans.
• A detailed analysis and insights into the effectiveness of using LLM-based role-play architectures for portraying believable character personalities.
• Actionable findings on the practical challenges and potential of integrating generative AI for conversation-based interactions to enhance user engagement and immersion in real-time virtual environments.

This Master's thesis delves into the application of Large Language Models (LLMs) for autonomous real-time decision-making within the Unity game engine. The primary focus is on engineering and evaluating an LLM-based architecture that enables game agents, either as dynamic non-player characters (NPCs) or as autonomous playtesting agents, to perceive their environment and execute complex, goal-oriented actions in real-time. This research adapts cutting-edge concepts from General Capable Agents to the practical, high-speed constraints of modern game environments, exploring a domain similar to that of projects like llm-colosseum.

Key Tasks

• Literature Review on LLM-driven Autonomous Agents: The student will conduct a comprehensive review of the current state-of-the-art in LLM-based agent architectures. This involves studying foundational concepts in reasoning-and-acting (ReAct) frameworks and agentic reflection. This review will provide the theoretical basis for designing an agent capable of operating effectively within a real-time game loop.

• Implementation of an LLM-based Agent Architecture for Unity: The core practical component involves designing and implementing a robust system to connect or integrate an LLM-based agent in a Unity environment (possibly identifying an existing open-source Unity game). The objective is to create an agent that can parse simplified game-state data, use the LLM to reason about a specific task, and translate the model's textual output back into concrete, executable actions within the game. This will demand Unity (C#) and Python skills, efficient handling of local LLMs, and prompt engineering to manage the agent's perception and decision-making process.

Expected Outcomes

• Understand the feasibility and limitations of using LLMs for high-frequency, real-time decision-making in a game, addressing the core research question.

• Implement a functional prototype of an autonomous agent within a Unity scene that demonstrates the ability to perceive its surroundings and independently complete a multi-step task.

• Critical insights into the primary challenges (e.g., LLM latency, state representation, action discretization) and the future potential of integrating LLM-based reasoning to create more engaging agents or more effective automated playtesting solutions.

This thesis represents an opportunity to contribute to the rapidly advancing field of AI in game development, offering significant practical experience in implementing sophisticated, language-based reasoning within a real-time, interactive application.

Virtual reality (VR) environments increasingly rely on the interaction between users and virtual agents, particularly in tasks requiring teamwork and collaboration. However, managing verbal communication with multiple agents presents unique challenges, such as discerning the intended recipient and ensuring accurate interpretation of commands. This thesis aims to address these challenges by investigating and developing systems that facilitate seamless communication with multiple virtual agents in VR settings.

The project will begin with an analysis of current solutions and approaches for multi-agent communication in VR. Based on these findings, a VR application will be developed to implement some of these systems for managing verbal interactions with multiple agents in a shared context. These systems may involve different strategies for agent recognition, dialogue management, and Natural Language Processing (NLP).

The final phase of the thesis will involve conducting user tests to evaluate the implemented systems. These tests will assess key factors such as usability, efficiency, and user satisfaction, providing insights into the strengths and weaknesses of each approach.

Required Skills:

• Unity development

• VR implementation techniques

• Basic understanding of Natural Language Processing (NLP)

In the following section, we present a series of thesis proposals, some of which necessitate management as internships due to constraints related to the access of equipment or facilities at the CRR.

It is important to note that the CRR has a limited capacity for hosting students, so project availability is subject to available spots. Any additional projects may start upon the completion of the theses already underway.

All the following theses are eventually available for multiple students working on the same project

Generative AI for data compression: a paradigm shift in multimedia transmission

This thesis explores a novel approach to data compression leveraging generative intelligence, particularly in the context of multimedia transmission. Traditional compression techniques focus on reducing file size while preserving fidelity, but generative models offer a paradigm shift: transmitting minimal data—such as a static image—alongside descriptive metadata or textual instructions, and reconstructing dynamic content (e.g., video or animation) at the destination using generative AI. This method significantly reduces bandwidth requirements by offloading complexity to the generative model, which synthesizes motion, context, and temporal changes from compact representations. The research investigates the efficiency, fidelity, and practical applications of this technique, including its potential in streaming, remote collaboration, and immersive media experiences.

https://arxiv.org/html/2405.20032v2

Detecting unauthorized AI training: copyright traps for generative AI

This thesis investigates the development and application of copyright traps designed to detect unauthorized use of creative content in the training of generative AI models. Building on research initiated at Imperial College London on large language models (LLMs), the work explores techniques for embedding hidden, traceable data within visual media—specifically images an videos—to serve as digital watermarks or forensic markers. These embedded signals allow content creators to later verify whether their work has been ingested by AI systems, offering a novel mechanism for intellectual property protection in the age of generative models. The study evaluates the robustness, detectability, and stealth of such traps, and proposes a framework for scalable implementation across various media formats.

Reference: https://www.imperial.ac.uk/news/255193/hidden-data-could-reveal-ai-model/

Quality assessment of AI-generated images and videos

This thesis presents a comprehensive evaluation of image quality produced by generative AI tools using objective metrics. As generative models become increasingly capable of producing photorealistic and artistic images, assessing their output quality is essential for applications in design, media, and scientific visualization. Sometimes the quality does not meet the high standards required by media companies particularly in terms of hallucination. The research applies no reference metrics such as BRISQUE, NIQE, e PIQE, and hallucination detection to benchmark various models, comparing their performance across different image types and use cases. The goal is to establish a standardized methodology for quantifying generative AI image\video fidelity and quality.

The ability to detect anomalies in driving scenes is crucial for the safe deployment of autonomous vehicles. Current benchmarks for anomaly segmentation, such as FishyScapes, SegmentMeIfYouCan (SMIYC), and RoadAnomaly, offer some insights into this challenge. These datasets primarily focus on anomalies as new, unseen objects placed on the road, represented through monocular camera images. However, this narrow definition fails to capture the broader and more complex nature of anomalies. Anomalies can also include familiar objects in abnormal configurations (e.g., a fallen tree or a car in an unusual position), as well as environmental or contextual irregularities. This thesis proposes to address these limitations by creating a new synthetic dataset using the CARLA driving simulator. In particular, we aim to:

• Support multi-modal anomaly detection methods, by collecting a dataset with multiple simulated sensors (RGB cameras, depth cameras, Lidar, radar, event cameras)

• Support the investigation of these methods across different illumination and weather conditions. This raises some challenges for the realism of the simulated sensors (e.g., implementing the effect of rain on the lidar sensor)

• Integrate both static and moving anomalies from a diverse set of categories (animals, objects)

• Include anomalies that are not just new object categories. For example, an anomaly could also be a common object category but in a strange configuration (e.g., a fallen tree or a car in an accident).

• The development of a well defined, customisable and possibly automatic pipeline to create short simulation episodes in CARLA, to streamline the data collection.

Substance use disorders (SUDs) are a critical global health challenge, affecting millions of individuals and their communities. Addressing the behavioral, psychological, and neurobiological dimensions of substance use requires innovative tools that can simulate real-world scenarios and offer controlled environments for study and intervention. VR technology has emerged as a powerful platform for creating immersive and interactive simulations, offering unique opportunities to explore and address substance use in a safe, replicable, and engaging manner.

This thesis focuses on the design, development, and testing of interactive 3D virtual reality environments to simulate real-world scenarios associated with substance use and misuse. The project will explore the creation of multiple immersive environments tailored to represent diverse settings, including social gatherings, clinical contexts, and educational spaces. These environments will serve as platforms for both research and intervention, enabling the study of substance use behaviors and the testing of innovative treatment strategies.

The proposed work aims to incorporate advanced technologies to enhance the functionality and interactivity of these virtual environments. This includes the integration of tools for recording and coding eye-tracking, motion data, and ecological momentary assessments (EMA), as well as leveraging LLMs to enable context-aware, dynamic interactions.

The main project activities include:

• design and development of multiple immersive, interactive virtual environments that accurately replicate real-world scenarios associated with substance use and its treatment.

• design VR-embedded surveys to collect data and assess user experiences.

• design processes to record, extract, and analyze various forms of interactive data, including eye-tracking, motion tracking, physiological responses, and subjective feedback.

• incorporate realistic avatars and life-like models to enhance user engagement and the authenticity of interactions.

• leverage large language models to enable dynamic and context-aware interactions, ensuring that virtual scenarios are responsive and adaptive to user behavior.

The project will be conducted in collaboration with the Colloca Lab at the University of Maryland, Baltimore, USA, and primarily based at the Polytechnic University of Turin. This partnership offers a unique opportunity for interdisciplinary research, combining expertise in VR technology, behavioral health, and clinical practice. Additionally, there is potential for a visiting period at Colloca Lab.

This thesis focuses on the integration of memory modules into large language model agents (LLM-Agents) to improve their ability to retain, retrieve, update and in general utilize information across interactions and from external knowledge sources. The primary objective is to design and develop memory architectures specifically tailored for LLM-Agents, enabling intelligent prompt management to integrate memories from past interactions or information from external knowledge seamlessly.

This research involves a comprehensive review of the state of the art, with a particular focus on open-source models to ensure resource efficiency and performance feasibility. Additionally, it aims to develop LLM-Agents optimized for specific applications and domains—such as AI companions, personalized learning platforms, and AI secretaries—while establishing robust evaluation methods to assess the performance of the memory modules and the LLM-Agent as a whole.

Based on the review, the LLM-agent developing will focus on the following tasks:

• Design and develop a memory architecture to integrate into LLMs, enabling efficient memory usage and external knowledge incorporation.

• Identify and leverage free and open-source models that ensure feasibility in terms of both resource requirements and performance.

• Develop an LLM-Agent tailored to a specific application and domain, such as AI companions, personalized learning platforms, or AI secretaries. Evaluation and Validation

• Identify the main methods and strategies for evaluating memory modules and the LLM-Agent as a whole.

• Develop and/or utilize datasets and benchmarks to support comprehensive evaluation.

• Test various models and architectures on the identified benchmarks, with a particular focus on understanding whether and how open-source models can support such applications effectively.

General Information about the theses proposals:

Duration: 6 or 12 months Expense reimbursement: €500–800 (to be evaluated) Work locations (flexible distribution):

• Varese (Buguggiate – Tecniplast headquarters, Gazzada/Schianno train station)
• STMicroelectronics Agrate (Colleoni Building)
• Work from home

Supervisor: Marco Garzola Application field: Preclinical research (experiments and final application testing are carried out in collaboration with our clients – universities and research centers)

Available opportunities:

• Master's thesis (curricular)
• Postgraduate internship (after master’s degree)

1) Title: Development of a Data Collection and Analysis Pipeline for Ultrasonic Vocalizations in Preclinical Rodent Studies

Description: This thesis focuses on supporting the collection of ultrasonic audio data in preclinical rodent experiments using the DVC 2.0 system. The candidate will perform a time-frequency analysis of the recorded data and develop an algorithm for the extraction of rodent vocalizations. By reviewing the existing literature, the candidate will cluster different vocalization types and implement a neural network for their recognition. The final objective is to create a real-time pipeline that processes live audio files, performs analysis, and extracts vocalizations, thus contributing to advancements in the field of rodent communication research.

2) Title: Development of a Tool for Detecting and Analyzing Rodent Interactions with Cage Enrichments Using Low-Resolution Fish-Eye Video Data

Description: This thesis focuses on developing a tool to detect recurring elements within a rodent cage, such as mouse houses and rolls, using low-resolution fish-eye video data collected with the DVC 2.0 system. The candidate will create and implement a machine learning algorithm to evaluate rodent interactions with these enrichment objects. The goal is to enhance the understanding of animal behavior in preclinical studies by providing a robust method for analyzing how rodents engage with their environment.

3) Title: 3D Reconstruction and Enhanced Tracking of Rodent Movement Using Synchronized Multi-Camera Video Data

Description: This thesis aims to synchronize VGA and fish-eye video recordings from multiple cameras in different positions, collected via the DVC 2.0 system, to achieve 3D reconstruction of rodent positions within the cage. The candidate will evaluate the movement of the animals in the 3D space and compare the performance of object detection models running on single-camera videos against a newly developed algorithm that utilizes object detection from both cameras to more accurately determine the animals' movements. The project will advance tracking methodologies in preclinical studies by enhancing spatial accuracy.

Training and learning in XR (VR/AR) in several scenarios (industrial, medical, educational) can be envisage as actvities characterized by a sequence of activities that can be organized in procedures. The activity organization can differ according to the specific scenario (e.g., activities can be sequential, alternative, looped and so on), anc can be generally represented as a graph of activities.

The learning phase is then usually organized in different steps (or phases):

• Learning: In a learning session, users are guided through the various actions the given procedure is organized into. Each action is introduced by visual and audio hints, aimed at explaining learners what they have to do, why, which are the objects they have to interact with and through which mechanics. Learners are then required to complete the activity in the XR environment, and the system controls the action execution and provide appropriate feedback to the learner

• Rehersal: where users can train the procedures they have learnt in the previous phase. Users can freely perform any action, but they cannot benefit from the visual and audio instructions available in the learning mode. However, the system can provide cognitive feedback (in terms of audio or other hints) to the user. Game/gamification mechanics can be included in this phase to help learners achieve the expected learning goals

• Evaluation: where learned skills are automatically assessed by the system. Users are required to perform the same procedure they have practiced before, without any feedback from the system, which is then capable to record execution times, correct and incorrect executions and any other problem arising during evaluation. Assessment is usually performed against a list of predefined metrics (which depends on the learning/training scenario)

As said before, this structure is standard in many application fields. The general objective of this proposal is to facilitate the development of such learning program and improve their effectiveness.

Research methods

An educational path can be structured through different learning methodologies, different assessment systems and activities organization (looped sequences, repeat only mistakes, and so on). However, the effectiveness of these approaches and the best combination of learning/assessment/organization methods is also related to the context where learning/training activities take place.

The objective of this thesis is to make a review of the state of the art to identify the most promising approaches, and to validate their effectiveness in different real-world contexts, in terms not only of knowledge and skill acquisition, but also of their retention over time

Topic 1: software farmework for rapid prototyping of MR-based learning environments

A first thesis topic is developing a software framework that allows a fast deployment of a learning program by defining the structure supporting activities, procedures and their scheduling, so that the designer of the educational intervention is only required to: i) create the assets to be used in the MR environment, ii) define the logic of the single activity, and iii) design the activity graph that define the procedures.

Students are required to implement the framework and develop at least two different use case scenario (in different fields, e.g., medical and industrial) to test it.

Topic 2: usability and UX of the learning environments

A second thesis topic is analyze the problem in terms of usability and User eXperience, i.e. analyze how to better deliver instructional/educational content in immersive AR/VR experiences, how to develop effective HCIs (in terms of input/output) and how to actively support users during their learning program (e.g., by adding AI-driven virtual instructors that can provide a natural and “face-to-face” support for the user).

Students are required to analyze the problem, proposed alternative solutions that can be implemented in the HCI, and validate them trough quantitative/qualitative assessment involving a panel of users. For this task, at least one use case scenario (in any possible field…) must be developed from scratch

Topic 3: development of an effective debriefing support

A third topic is the development a debriefing companion application, which relies on the analytics (and other data) collected during the rehearsal and evaluation sessions. The availability of a debriefing step is extremely relevant for knowledge retention, since it helps learners to reflect on what they did, get insights from their experience and make meaningful connections with the real world, thus enhancing transfer of knowledge and skills. Even when results are not as successful as the learners hoped, debriefing can still promote active learning by helping them to analyze mistakes made and explore alternative solutions.

Students are required to analyze the problem, proposed alternative solutions that can be implemented in the debriefing companion app, and validate them trough quantitative/qualitative assessment involving a panel of users. For this task, at least one use case scenario (in any possible field…) must be developed from scratch.

The goal of this work is to evaluate and implement solutions that allow simultaneous access to a three-dimensional environment in which two or more users interact with each other and with the environment. Three application scenarios are proposed below:

1. Use for educational purposes in a classroom where the professor and students access the same application from different devices. A plausible scenario consists of:
2. Professor using a tablet or PC version of the application to highlight objects/points of interest;
3. VR device used by one or more students to perform operations on three-dimensional objects within the scene;
4. Touch device (e.g. lim) that allows interacting with objects in the scene and performing the same operations intended for virtual reality devices, but with different input devices (touch screen);
5. Use for remote assistance, which allows a technician/student (equipped with smart glasses) who wants to perform work on a machine/a trainign activity to receive real-time assistance from a senior operator/an expert who can connect to the machine via an app (desktop or VR) and view its 3D and related data. The senior can geographically locate the junior using the 3D model or the camera on the smart glasses and tell him where to go to work. It would be interesting to be able to give the junior the feeling of the senior's presence as well.

The objective of the thesis will be to evaluate the effectiveness of the possible solutions through the implementation of different use cases in both medical and industrial scenarios.

The Laboratory

Rehab Technologies Lab is an innovation lab jointly created by IIT and INAIL (National Institute for Insurance against Accidents at Work) to develop new high tech rehabilitation solutions of high social impact and market potential. The projects so far include: the CE marked poly-articulated hand prosthesis (Hannes), the upper-limb (Float) and the lower-limb (TWIN) exoskeletons, both developed in compliance with the ISO standards for medical devices. Moreover, the laboratory leads and participates in neuro-rehabilitation projects aimed at studying cognitive and physical workload to help neurological patients in improving the quality of life.

Motivations and general objectives

The proposed activities are within the framework of the “NRTWIN - NeuroRobotic TWINning” project, aimed to design, develop, and test a set of neuro-robotic solutions (sensors, computational models, control systems) for replicate virtually physio-motor activities to study mental and physical effort for different population of subjects: healthy individuals, prosthetic users, people with Multiple Sclerosis (MS).

In general, the analysis and understanding of the pick-and-place task is fundamental in various industrial and rehabilitation settings, involving the movement of objects from one location to another. Improving the performance of this task and the required motion cost can have significant implications for ergonomics, productivity, and physical rehabilitation. Virtual Reality (VR) offers a unique platform for creating immersive and controlled environments to study and enhance motor skills. This thesis aims to contribute to the development of a VR-based pick-and-place exergame and to collect and analyze game data to understand the interaction of the individuals with the virtual environment through motion tracking data related to the execution of the task.

Required skills

The exergame is developed inside Unity engine and C#. The candidate is required to expand the existing game structure inside the engine and create customized scripts for data recording (C#) and analysis (MATLAB or Python).

Working location

The candidate can choose between working and being supervised remotely or working on site at the Rehab Technologies Laboratory, IIT (Morego, Genoa, Italy). The remote option will require some short stays (within a week) in Genoa for data acquisitions or application testing.

Max number of students: 1

This thesis focuses on the design and development of computer vision algorithms for pose estimation of space objects, a critical task for enhancing in-orbit services and space sustainability. The project involves creating and validating these algorithms using both synthetic data, generated through tools like Blender, and realistic data collected from a test bench. The research aims to contribute to the advancement of technologies that support safe and efficient space operations, aligning with the mission of Astroscale to promote long-term orbital sustainability.

Main Missions

• Design and development of Computer Vision algorithms for pose estimation of space objects

• Validation of the algorithms using synthetic data, as well as realistic data acquired from a test bench.

Essential Skills

• Currently enrolled in a general or specialized engineering school in the computer vision and machine learning field.

• Experience with synthetic generating images software (i.e., Blender, SurRender)

• Proficiency in Python and deep learning computer vision frameworks (i.e., TensorFlow, PyTorch)

• Willingness and ability to work in a multidisciplinary and international environment.

• Innovative and proactive mindset.

• Excellent communication skills.

• Fluent in French and English.

Desired Skills

• Experience in aerospace projects.

About Astroscale

Founded in 2013, Astroscale is the first private company with a mission to secure long-term spaceflight safety and orbital sustainability for the benefit of future generations. Astroscale develops in-orbit services that improve the characterization and deorbit of space debris or extend the life of old satellites. Headquartered in Japan, Astroscale has offices in the United Kingdom, United States, Israel and France. Astroscale has launched its pioneering technology missions, ELSA-d in 2020 and ADRAS-J in 2024, achieving the world’s first successful approach to a non-cooperative space object. Astroscale France (ASFR) was founded in mid-2023 and has grown to approximately 20 staff. Located in the central quarter of Toulouse, we are developing cutting-edge on-orbit services projects, such as a multi-target inspection mission.

Contact

v.pesce@astroscale.com c.magueur@astroscale.com

The aim of this thesis is to create an advanced digital twin of an urban environment, specifically focusing on a neighborhood in Turin, Italy, and integrating it with the CARLA simulator. This project encompasses the development and integration of various components required for a highly effective and realistic driving simulator.

Goals and Objectives:

Urban Environment Replication:

• Accurately replicate roads, buildings, and landmarks of a specific neighborhood in Turin within the CARLA simulator.

• Ensure the digital twin includes detailed environmental features to enhance realism.

Integration of Autonomous Agents and Realistic Elements:

• Integrate autonomous agents such as vehicles, and vulnerable road users, ensuring their behavior mimics real-world scenarios.

• Simulate realistic traffic flow and patterns to mirror actual urban driving conditions.

• Incorporate dynamic weather conditions to test vehicle performance under varying environmental factors.

Real-Time VR Environment Implementation:

• Enable the use of this digital twin within a real-time VR environment for immersive simulation experiences.

• Focus on achieving high performance and low latency to ensure a seamless VR experience.

Methodology for Continuous Update and Replication:

• Develop a methodology that allows for continuous updating and modification of the digital twin.

• Create a framework that can be replicated for different urban and rural contexts, enhancing the versatility of the project.

Use of Advanced Technologies:

• The project should be implemented with Unreal Engine 5 for its cutting-edge graphical capabilities, ensuring a visually realistic simulation.

• The implementation will leverage CARLA’s open-source framework for its modularity, facilitating integration with various simulators and programs.

Research-Oriented Tool Development:

• Aim to create a tool that can be used for research purposes, particularly in fields like autonomous driving, urban planning, and traffic management.

• Ensure that the tool is adaptable for future technological advancements and research needs.

This thesis represents a blend of simulation technology, urban planning, and software engineering. It offers a unique opportunity for students to contribute to the growing field of digital twins, particularly in the context of urban environments and autonomous driving simulation. The project not only aims to create a realistic digital replica of a neighborhood but also establishes a replicable framework that can be applied to various urban settings, thereby contributing significantly to research and development in this field.

This thesis deals with the development of a tool for the development of sophisticated and modular Head-Up-Displays (HUD) for the CARLA driving simulator. HUDs, which project important information onto a vehicle's windshield, have become increasingly prevalent in modern vehicles and offer the advantage of reducing driver distraction and increasing road safety. This tool aims to replicate and extend the functionality of real HUDs in a virtual driving environment.

Aims and objectives:

Development of a modular HUD tool:

• Create a HUD tool that allows for extensive customization, including adjustable dimensions, transparency levels, and color schemes.

• Ensure that the tool is capable of implementing various interfaces and systems, such as navigation aids, warning systems, safety distance indicators and support for Non-Driving Related Tasks (NDRT).

Integration into the CARLA simulator:

• Develop the HUD tool using Unreal Engine 5 to ensure seamless integration with the CARLA driving simulator, a widely recognized open-source platform for autonomous driving research.

• Focus on compatibility and optimization for Virtual Reality (VR) to enhance the immersive experience of the driving simulator.

Diverse interface and system testing:

• Enable the HUD tool to support a range of test cases, such as different driving scenarios, environmental conditions and vehicle types.

• Integrate functions for the simulation of advanced driver assistance systems (ADAS) and autonomous driving technologies.

Design the user interface and driving experience:

• Emphasize a user-friendly HUD tool design that allows users to easily create and modify HUD elements. Implement intuitive controls and visualization techniques to enable efficient design processes.

Evaluation through case studies:

• Conduct a detailed case study to evaluate the effectiveness, usability and potential of the developed HUD tool.

• Analyze the impact of different HUD designs on driver behavior, attention and overall driving experience in the simulator.

Contribution to research:

• Contribute to research in the field of driving simulation and automotive user interface

• Explore the potential of HUDs to improve driver awareness and safety, especially in the context of autonomous vehicles and advanced driving simulators.

PRODUCTION OF A GENERATIVE BOOK ON THE CLIMATE AND ENERGY TRANSITIONS IN THE MEDITERRANEAN AREA APPLYING A LARGE LANGUAGE MODEL (LLM)

The context: EST has produced in the last years five (5) reports on the climate and energy transition in the Mediterranean area, accumulating a large body of knowledge, data and references. The reports gather information on all the energy technologies (from hydrocarbons to renewables), on maritime transport, and on their emissions of greenhouse gases.

The problem: to structure that volume of information in a manner that is ready and fit for all end users, enabling assisted interactions (open and with predefined prompts), and that can grow with the addition of new information. This goal derives from the limitation of traditional books where contents are static and too profuse, and therefore they are on the one hand quickly outdated, and on the other difficult to consult and not prone to rapid answers to the requests of the reader.

The thesis activity to develop a so-called Generative Book (GB) using LLM technologies, applying the platform BLOOM. The new GB will, using as the basic content the already available five reports and related sources, enable users to interact with the contents in different ways (quick summaries, put forward different questions, develop questionnaires based on the text, obtaining usable output against specific requests, etc); will let users contribute to the contents by for instance indicating useful sources of information, indicating shortcomings, commenting, etc.; and will act as an evolving platform to facilitate the growth and evolution of the contents.

CUSTOM TRAINING AND PROMPT ENGINEERING OF A LLM PLATFORM ON THE ENERGY AND CLIMATE TRANSITIONS IN THE MEDITERRANEAN AREA

The context: EST has produced in the last years five (5) reports on the climate and energy transition in the Mediterranean area, accumulating a large body of knowledge, data and references. The reports make reference to a vast set of documents and data. EST intends to upload all those elements in an LLM-based Generative Book (GB).

The problem: to accelerate the adaptation of an out-of-the-box LLM (BLOOM) for its use with knowledge and contents referring to the Mediterranean energy and climate area.

The thesis activity: to compose and format prompts to maximize the model’s performance regarding the tasks defined for the GB, and to custom train the GB model with datasets taken from the EST reports. This will involve fine-tuning the training parameters, setting up the training environment, and fine-tuning the GB model.

INTEGRATION OF JUPYTER NOTEBOOKS WITH A LLM PLATFORM ON THE ENERGY AND CLIMATE TRANSITIONS IN THE MEDITERRANEAN AREA

The context: EST has produced in the last years five (5) reports on the climate and energy transition in the Mediterranean area, accumulating a large body of knowledge, data and references. Many values are supported by equations and formulae, which are not made explicit in the reports.

The problem: to produce an interactive environment composed of computational documents using the data, equations and explanations present in the EST reports on the energy and climate in the Mediterranean ready for their customised use, visualization and analysis, and integrated into a Generative Book (GB).

The thesis activity: to produce Jupyter notebooks concerning energy and climate in the Mediterranean area to be integrated in an LLM-based GB, connecting software codes, data analytics and text, to work interactively and being customizable by the end users.

DEVELOPMENT OF A DIGITAL PLATFORM FOR SUPPORTING TABLE-TOP EXERCISES APPLIED TO THE CLIMATE AND ENERGY TRANSITIONS IN CITIES

The context: EST supports cities in the elaboration of: i) their transition towards climate neutrality and the related production of Climate City Contracts, and ii) plans for their sustainability. For these goals, EST is developing two digital platforms, CLICC and CITTA, composed of interactive tools for dealing with data and text in a multimedia environment, and a full set of scientific instruments for the calculation and analysis of data. The study of future scenarios for climate neutrality and sustainability demands the interaction with all stakeholders, and the joint study of best alternatives concerning all potential contingencies. These interactions can be structured in the form of Table Top Exercises (TTXs).

The problem: to facilitate the arrangements and implementation of TTXs for cities by means of digital applications in an interactive environment, with the management of narrative scripts, a diversity of timing scales, alternative paths in the presence of contingencies, etc., while recording the decision and actions of all participants. TTxs are role-playing activities in which players respond to scenarios presented by the facilitators.

The thesis activity: to produce an interactive digital platform using open-source technologies for supporting the preparation and the running of TTXs applied to the climate neutrality and sustainability of cities, taking advantage of the existing platforms CLICC and CITTA. The platform should provide facilities for ex-ante preparation of the TTX, for the work of the participants (i.e. players, observers, facilitators, note takers), and for the ex-post analysis and reporting.

DESIGN OF AN INTERACTIVE INTERFACE FOR THE DIGITAL TWIN OF CITIES FOR THE STUDY OF CLIMATE NEUTRALITY AND SUSTAINABILITY

The context: EST supports the city of Torino in their plans to climate neutrality and sustainability. A crucial aspect of this support is to enable the city administrators and all city stakeholders to visualize and interact with a digital twin of the various main components of the city (e.g. energy, transport, waste, green areas, etc.) directly related to the production and mitigation of emissions. EST is developing two digital platforms, CLICC and CITTA, composed of interactive tools for dealing with data and text in a multimedia environment, and a full set of scientific instruments for the calculation and analysis of data. EST operates a Decision Theatre with a 180 degrees, 3 meter tall wall where to display interactive software applications.

The problem: to facilitate the interaction with the manifold aspects related to climate neutrality and sustainability of cities, including the virtual representation of the city systems as a digital twin. This representation should include both past data and future scenarios, with the possibility of displaying the evolution in time of those scenarios.

The thesis activity: to produce an interactive interface based on open-source tools such as Unity, to be used in both EST’s Decision Theatre and the web, able to dynamically exhibit data in 2D/3D, and create game-like experiences based on the climate neutrality and sustainability scenarios produced by CLICC and CITTA. The activity will be applied to the city of Torino.

Technological advancements have made it possible to use interactive virtual scenarios to simulate reality using advanced rendering and graphic techniques. These scenarios can be used in various contexts, such as product presentation and personnel training in the management of industrial machinery. However, graphic optimization for different devices is crucial to ensure a smooth and high-quality experience.

The objective of this thesis is to analyze the use of interactive virtual scenarios in personnel training and product presentation, with particular attention to graphic optimization for different devices. The expected results includes highlighting the advantages and analyzing the challenges associated with the current proposal. The results obtained can be used by companies to improve user experience and ensure a good quality of interactive virtual scenarios.

Required skills Basic skills in the field of 3D graphics, software development and game engine programming.

The activities will take place in Turin:

SynArea Consultants C.so Tortona 17

Polythecnic of Turin (when possible)