| KEYWORD |
Area Engineering
Study and development of a conversational support agent designed for automatic speech detection in a single-pilot civil aviation scenario.
Reference persons ELISA CAPELLO
External reference persons Mario NIGRA (SIPAL), Nicola PUCA (POLiTO)
Research Groups 08- Dinamica, controllo e simulazione del volo
Description Context of application: As human errors are still considered a major factor in aircraft accidents, the introduction of automated agents can offer support to operations in the cockpit. Currently, efforts are underway to evaluate new intelligent subsystems to assist crews in the safe management of the flight or possibly meet the emerging requirements associated with the Single-Pilot Operations (SPO) paradigm. Although the potential benefits, SPOs are still debated since the same or even higher level of safety should be demonstrated with respect to normal dual-crew operations. Currently proposed forms of Virtual Pilot Assistants (VPAs) are based on conditioning the on-board level of automation according to pilots’ behavioral deviations from a predetermined workflow. Given the absence of the co-pilot, the assignment of tasks, roles, and responsibilities to automation becomes critical, as the goal would be to effectively reduce the operator's physical and cognitive workload.
One more approach is to set-up Artificial Intelligence (AI)-based aiding subsystems to enable pilots to focus more on decision-making processes and support pilots with multiple interaction modalities. To this extent, research into Automatic Speech Recognition (ASR) has been emerging as a crucial step element for Single-Pilot Operations. Voice can be in fact one of the most direct and resource-saving hints the Pilot Flying (PF) would pursue to interact with a non-human flight deck. Chatbots powered by Large Language Models (LLMs) represent a central technology for that, having the ability to answer users’ questions in various contexts. On the other hand, their effectiveness strongly relies on domain-specific knowledge.
Aim of Thesis: Study and development of a conversational support agent designed for automatic speech detection in a single-pilot civil aviation scenario. Our objective is for the student to customize an existing Natural Language Processing (NLP) solution to effectively manage tasks such as question answering or pilot intent detection during specific flight segments. Apart from overload, an important aspect to consider for the transition to SPOs is to keep the pilot involved during undemanding flight phases, too. One potential application for an assistance audio interface, therefore, can be to handle some of the communication tasks (e.g. channels selection, hints for procedures) or perform cross-checking activities (e.g. checklists, briefings) where the single pilot level of attention needs to be triggered. All the thesis activity will be conducted in collaboration with SIPAL S.p.A.
Required skills Analisi di missione, linguaggi di programmazione
Notes The candidate is expected to fine-tune a general-purpose language model through the augmentation with domain-specific external knowledge, evaluating the pros and cons of supervised and unsupervised adaptation techniques. A prevalent challenge with Large Language Models (LLMs) is the generation of incorrect information, often referred to as "hallucinations”, especially when queries extend beyond the model’s training data or necessitate up-to-date information. Such a characteristic poses a potential hazard in the fast-paced and demanding aircraft environment, where decisions must be optimized promptly considering time constraints, resources, as well mission goals. One promising approach to address these limitations is Retrieval-Augmented Generation (RAG), which leverages the power of existing models and external data (e.g. private documents) to enhance the responses generation process. According to the RAG paradigm, LLMs can be prompted to answer based on a context-enriched prompt. Specifically, the student will be tasked with reflecting the intricacies of civilian aircraft cockpit operations within a document corpus aimed at the external dataset creation. Thus, a good knowledge of fundamental cockpit operations is required. At the conclusion of the thesis, answer generation to domain-specific questions must be proved (the scenario of a selected flight mission can be involved). At the same time, the student can devise strategies and requirements for implementing modifications within the flight deck environment, if necessary. This entails a careful analysis of operational needs, safety considerations, and the potential impact on overall cockpit functionality.
Deadline 10/07/2025
PROPONI LA TUA CANDIDATURA