GuardAD: Safeguarding Autonomous Driving MLLMs via Markovian Safety Logic
Researchers have introduced GuardAD, a novel model-agnostic safeguard designed to enhance the safety of Multimodal Large Language Models (MLLMs) in autonomous driving systems. Addressing the limitations of static logical constraints that fail to account for dynamic traffic interactions, GuardAD formulates safety as an evolving Markovian logical state. It utilizes Neuro-Symbolic Logic Formalization and n-th order Markovian Logic Induction to infer emerging and latent hazards beyond single-step observations. Unlike traditional methods that simply veto unsafe actions, GuardAD employs Logic-Driven Action Revision to actively refine driving decisions without modifying the underlying MLLM. Extensive experiments across multiple benchmarks demonstrate that this approach significantly reduces accident rates by 32.07% while improving task performance by 6.85%. The effectiveness of GuardAD has been further validated through closed-loop simulations and physical-world vehicle studies, highlighting its potential to robustly protect autonomous vehicles in accident-prone scenarios by integrating temporally grounded safety reasoning into dynamic driving environments.
Wire timeline
GuardAD: Safeguarding Autonomous Driving MLLMs via Markovian Safety Logic
Researchers have introduced GuardAD, a novel model-agnostic safeguard designed to enhance the safety of Multimodal Large Language Models (MLLMs) in autonomous driving systems. Addressing the limitations of static logical constraints that fail to account for dynamic traffic interactions, GuardAD formulates safety as an evolving Markovian logical state. It utilizes Neuro-Symbolic Logic Formalization and n-th order Markovian Logic Induction to infer emerging and latent hazards beyond single-step observations. Unlike traditional methods that simply veto unsafe actions, GuardAD employs Logic-Driven Action Revision to actively refine driving decisions without modifying the underlying MLLM. Extensive experiments across multiple benchmarks demonstrate that this approach significantly reduces accident rates by 32.07% while improving task performance by 6.85%. The effectiveness of GuardAD has been further validated through closed-loop simulations and physical-world vehicle studies, highlighting its potential to robustly protect autonomous vehicles in accident-prone scenarios by integrating temporally grounded safety reasoning into dynamic driving environments.
cs.AI updates on arXiv.org