A Runtime Cognitive Diagnosis Framework for Autonomous Vehicles

Autonomous vehicles (AVs) are transforming transportation, evolving from research prototypes to operational services worldwide. However, they continue to face challenges due to the complexity and unpredictability of real-world environments, with negative news about AV safety capturing headlines. A tragic example is the Xiaomi SU7 incident on March 29, 2025, which resulted in three fatalities, highlighting the ongoing journey toward ensuring AV safety.Safety diagnosis, which assesses whether an AV can make safe driving decisions across various traffic scenarios, is essential. Existing efforts are limited in two key ways. First, most rely on offline testing to identify faults before deployment, yet vehicles will inevitably encounter new situations while operating on roads. Second, although some runtime diagnosis solutions exist, they are designed for AV architectures with multiple discrete modules and aim to identify which module caused failures. These approaches have become outdated as AVs evolve toward end-to-end architectures, where multiple modules are consolidated into a single monolithic neural network—analogous to the human brain—posing significant challenges for diagnosis.In this project, we propose the first runtime diagnosis framework that monitors end-to-end AVs in real time to predict their safe operational capability. If safety cannot be ensured, it diagnoses the underlying causes for timely assistance. Inspired by cognitive science, our approach treats an AV similarly to how a doctor examines a human patient, redefining diagnosis from a module-level to a cognitive function-level. Preliminary experiments reveal that end-to-end AVs exhibit failures in perception, memory, and reasoning, akin to human cognitive disorders ADHD, amnesia, and schizophrenia. Accordingly, we will develop three diagnostic mechanisms: Task 1 implements a meta-attention-like mechanism to examine the AV model's perception status, ensuring accurate comprehension of the traffic environment; Task 2 introduces a meta-memory-like mechanism to monitor the model's memory boundary, preventing overconfident decisions based on unfamiliar knowledge; Task 3, inspired by reality testing, inspects the model's reasoning capabilities, ensuring it operates within a range resilient to natural and adversarial input perturbations.Together, these tasks create a comprehensive runtime safety diagnosis framework that integrates cognitive science, deep learning, and AV systems. Our solutions will be tested both in the lab and in real autonomous vehicles, ensuring seamless industry implementation and maximizing the real-world impact of our research. Centered around a fundamental perception-memory-reasoning triplet, this framework can immediately enhance AV driving safety while holding significant potential as a universal diagnostic tool for diverse AI and autonomous domains.