vail-ucla/navsafe-infty

NavSafe benchmarks closed-loop driving-policy safety in reconstructed scenes. A user supplies scenario data, a policy adapter and checkpoint, then evaluation configures the simulator and renderer, obtains policy observations and actions, executes the ego alongside replay/traffic actors, and records scores and artifacts. The repository also contains log-ingestion and parsing paths for preparing scenarios, model implementations, sensor/render support, and visualization. The README documents scenario mining, reconstruction preparation, asset harvesting, and recipe authoring; their implementation directories are not present in the supplied tree, so those branches are not expanded here. Some runtime links remain uncertain because evaluator and environment implementations were not excerpted.