Ensuring Autonomous Vehicle Safety in Cities

Summary & Methodology Analysis

The framework addresses the challenge of autonomous systems in unstructured urban environments by moving away from static safety constraints. Instead, it utilizes a hierarchical world model that separates concerns into three layers: a symbolic layer for traffic rules and policies, a spatial layer for occupancy and geometric motion planning, and a dynamic layer for physical actuator constraints. This architecture enables bidirectional information flow where safety requirements propagate downward while feedback loops allow for adaptation based on dynamic conditions. The methodology relies on generating synthetic training scenarios specifically derived from explicit assets, exposure modes, and hazard scenarios to establish a learned urban safety envelope. By utilizing a cross-layer transformation process, the system maintains safety as a holistic property rather than an isolated module check. The system is managed by a Physical AI Harness that performs three primary operations: Projection, which forces AI outputs to remain within the defined safety envelope, Isolation, which allocates dedicated resources to AI workloads, and Transfer, which hands off control to a fallback controller if necessary. While this design provides a structured approach to safety, the framework is ultimately limited by the quality and completeness of the hazard analysis used to define the safety envelope. Consequently, scenarios not accounted for in the initial hazard analysis may still pose challenges to the system's safety. Validation for this proof-of-concept is currently underway using the NVIDIA Isaac Sim, Isaac Lab, NVIDIA Cosmos, and NVIDIA Omniverse stack.