The Facets of Edge AI in Automotive: Exploring Embedded Frameworks, Voice Assistants, and Deep Reinforcement Learning

This thesis presents a comprehensive exploration of the synergistic relationship between edge computing and AI, with a particular focus on the automotive sector. As technology rapidly evolves, edge computing emerges as a paradigm shift, especially significant in automotive applications. The shift from cloud-centric approaches to decentralized computation enhances real-time processing capabilities and reduces latency, enabling intelligent decision-making at the network's edge. Central to this investigation is Edgine, a versatile, non-vendor-locked framework tailored for heterogeneous IoT applications. This work not only evaluates Edgine's adaptability but also innovatively applies it in developing tools for performance assessment in the automotive industry. Indeed, Edgine proved useful even in most of the later developed tools, both as an evaluation and measurement tool. Another significant contribution of this research pertaining to edge AI vehicular technology is the development of embedded voice assistants optimized for vehicles. The thesis details the creation of an end-to-end voice assistant system capable of operating offline, emphasizing privacy and security concerns inherent in cloud-based systems. The system supports the Italian language and, in a vehicular-only context, can achieve results comparable to cloud-connected solutions, demonstrating the feasibility of integrating advanced AI methods in embedded systems and the application of transfer learning. The potential of edge computing in overcoming the limitations of traditional cloud-connected solutions is also examined, alongside future research directions for enhancing voice assistants in terms of latency, language, and domain support. The thesis then shifts focus to deep reinforcement learning (DRL), specifically its application to automated driving for low-speed maneuvering. The effectiveness of DRL is explored through experiments in both Unity and CARLA -simulated environments. Key factors for successful DRL training, such as curriculum learning and simulation parameter tuning, are discussed. Results, in both environments, are promising, paving the way to possible future research directions in dynamic scenarios and real-world vehicle implementations. The final area of exploration is the explainability of DRL models, a critical aspect in domains like automated driving where safety is paramount. A novel approach for interpretability analysis is presented, combining episode timelines, frame-by-frame analysis, and aggregated statistical analysis. This investigation provides insights into the decision-making processes of DRL models and highlights future research opportunities in areas such as temporal correlations and more complex vehicular models. In summary, this thesis links advancements in edge computing, embedded voice assistants, DRL in automated driving, and DRL model explainability. This integration shapes a dynamic and evolving landscape, fostering a foundation for innovative developments within the automotive industry.