Reproduce “Exploring the Potential of World Models for Anomaly Detection in Autonomous Driving” on OpenResearch

About this paper

Daniel Bogdoll, Lukas Bosch, Tim Joseph, Helen Gremmelmaier, Yitian Yang, J. Marius Zöllner

Researchers at FZI and KIT propose a conceptual framework demonstrating how a single world model can unify various anomaly detection methods (reconstructive, generative, predictive, confidence, feature-based) for autonomous driving. This approach leverages the world model's intrinsic components and advocates for controlled simulation data to robustly define "normality" for detection.

Key insights

• A single world model inherently possesses the functionalities required for all major anomaly detection categories, allowing for a unified and potentially more efficient system compared to multiple specialized models.
• Rigorous definition of "normality" through controlled training data, ideally from simulation, is crucial for effective anomaly detection and evaluation in autonomous driving.
• Integrating planned actions from the vehicle's planning module enables world models to generate realistic, action-conditioned future predictions, which are vital for anomaly detection in dynamic autonomous driving scenarios.

Summary by alphaXiv.