Trustworthy Autonomous Vehicles Architecture

Autonomous Vehicles (AVs) rely heavily on Artificial Intelligence (AI) and Machine Learning (ML) to navigate roads, recognize obstacles, and make driving decisions. However, a major issue with current AV architectures is that AI-based decision-making is often too slow or too complex for real-time safety-critical scenarios. Many AVs fail to react in time because their AI systems prioritize high-precision object recognition over immediate safety measures, leading to fatal crashes. Unfortunately, AV architectures from companies like Tesla, Cruise, Waymo, and Udacity have faced growing criticism due to increasing fatal incidents, largely attributed to failures in their AI/ML systems. For example, some AVs have detected obstacles seconds before an accident but took too long to classify them (e.g., identifying a pedestrian as a vehicle, then a bicycle, before realizing too late that braking was needed). Others completely failed to react because their AI systems were still processing the best possible decision while time ran out. These failures highlight a fundamental flaw: existing AV architectures do not properly balance AI intelligence with trustworthy, real-time safety mechanisms. In our paper [Savvy], we introduced Savvy, a novel architecture for Trustworthy Autonomous Vehicles that balances AI-driven intelligence with safety and reliability. The core problem it addresses is the failure of existing AV architectures to reconcile AI/ML-based stochastic decision-making with the deterministic, time-sensitive nature of driving control. Savvy architecture addresses this problem by introducing a concept called Time-aware Predictive Quality Degradation (TPQD), leveraging the tuning properties of tunable AI models like DNNs. Instead of waiting for highly accurate AI recognition, Savvy prioritizes timely and "good-enough" decisions within safety-critical time limits. This means that an AV using Savvy would not waste valuable seconds trying to determine if an obstacle is a deer or a dog— it would immediately classify it as a moving hazard and take necessary action (e.g., braking or steering away). In this project, we are studying the feasibility of current/new AI models for TPQD, the exploration and incorporation for sensing data (Radar, LIDAR, Cameras, etc.) to improve perception, improving the architecture details, define AI model for time estimation and trigger, and implement/evaluate the solution for simulation or integration with a robot car (which may need embedded background). The intern can help with any of these tasks. [Savvy] Shoker, Ali, Rehana Yasmin, and Paulo Esteves-Verissimo. "Savvy: Trustworthy autonomous vehicles architecture." arXiv preprint arXiv:2402.14580 (2024).

The objectives of the project will be to: - explore state-of-the-art research and practice together with PhDs and researchers; - help on defining and solving a problem conceptually; - implement a Proof-of-Concept solution with evaluation/simulation; and - contribute and coauthor a scientific paper. More details can be shared and defined after admission and specific topic selection.