Tesi - Addestramento tramite tecniche di Reinforcement Learning di entità virtuali generate al computer all'interno di un simulatore di volo

Questa posizione è in Leonardo

Riassunto dell'opportunità da parte della Joinrs AI: Leonardo cerca un neolaureato in Ingegneria Informatica o Informatica per uno stage finalizzato allo sviluppo della tesi sul controllo intelligente di velivoli sintetici mediante tecniche di Reinforcement Learning in ambiente simulato. Il ruolo prevede analisi, progettazione e validazione di modelli di AI per il controllo di volo. Il tirocinio ha durata massima di sei mesi con rimborso spese e si svolge presso la sede di Torino.

Il processo di selezione sarà interamente gestito da Leonardo.

--

# Job Description:

Leonardo is an international industrial group, one of the world leaders in Aerospace, Defense, and Security, which creates multi-domain technological capabilities in Helicopters, Aircraft, Aerostructures, Electronics, Cyber Security, and Space. With over 60,000 employees worldwide, the company has a solid industrial presence in Italy, the UK, Poland, and the United States, and operates in 150 countries also through subsidiaries, joint ventures, and shareholdings. A key player in major global strategic programs, it is a technological and industrial partner for Governments, Defense Administrations, Institutions, and businesses. In 2024, Leonardo recorded consolidated revenues of €17.8 billion, new orders of €20.9 billion, and invested €2.5 billion in R&D activities. Innovation, continuous research, digital industry, and sustainability are the pillars of its global business.

Experienced tutors in their sector will guide you, allowing you to deepen your theoretical knowledge and develop your thesis, preparing you in the best possible way for future professional challenges.

The topics proposed for theses to be developed at Leonardo cover a vast spectrum of technological, research, and innovation fields: from Artificial Intelligence to High-Performance Computing, from Cyber Security to Materials Engineering, through to the aerospace sectors. You will be able to explore the most cutting-edge areas of your field of study, with creativity and a spirit of innovation.

To support you during this experience, which will last a maximum of six months, an expense reimbursement is also provided.

We are looking for 1 young student to join us for an internship with the objective of developing their thesis on the topic of "Training virtual entities generated by computer using Reinforcement Learning techniques within a flight simulator" at the Turin site.

The purpose of this thesis is the study, complemented by the development, of flight control software for a synthetic aircraft based on an artificial intelligence model trained with Reinforcement Learning techniques. Flight control is usually dominated by PID (Proportional-Integral-Derivative) logics, and although effective, these systems have limitations that Reinforcement Learning (RL) aims to overcome.

An RL model has the ability to learn directly from interaction with the environment and to build an implicit representation of these dynamics, succeeding in capturing complex behaviors. Unlike the fixed parameters of a PID, an RL agent can learn a control policy capable of managing a wider variety of operating conditions. Thanks to training in simulated environments that include variable scenarios, the agent can develop more robust control techniques, improving the stability and performance of the system.

While a PID focuses on reducing an error (e.g., maintaining altitude), RL can be trained to maximize a complex reward function that integrates different performance criteria within the same learning process. In this way, the agent can be trained to find the best control policies that simultaneously balance different operational factors, such as precision in following a desired trajectory, fuel saving, and structural integrity.

The main objective of this thesis is the design and implementation of an intelligent agent capable of managing the dynamic control of a synthetic aircraft.

The work will be articulated in the following phases:

• Analysis of commercial products and integration: Study of flight simulators available on the market (e.g., X-Plane 12 or DCS World) as physics engines. The AI agent will interact with the simulator via APIs or dedicated plugins, receiving as input the aircraft's state (attitude, speed, position) and providing as output commands to the control surfaces;
• Definition of the Learning Model: Design of the "reward function" to train the agent to perform specific maneuvers (e.g., maintaining level flight, intercepting waypoints, or evasive maneuvers) by maximizing the parameters chosen for the project;
• Validation i