TRACE: Topology-aware Reconstruction of Accidents in CARLA for AV Evaluation
TRACE automates crash reconstructions using real-world data for AV testing.
TRACE, created by Nahian Salsabil and Sebastian Elbaum, revolutionizes the validation of Autonomous Vehicles (AVs) by automating the reconstruction of accident scenarios based on NHTSA crash reports. Utilizing a pipeline that retrieves specific OpenStreetMap data, TRACE ensures that the topological complexities of real-world crashes are accurately represented in CARLA simulations. By leveraging Large Language Models, TRACE infers vehicles' initial states from detailed road geometries and pre-crash maneuvers, leading to more realistic testing environments for AVs.
The TRACE pipeline generates a benchmark comprising 52 diverse accident scenarios, encompassing various collision types and road topologies. This open-source resource provides a critical tool for developers, enabling comprehensive evaluation of AV systems against real-world failures. With the ability to simulate rare, safety-critical scenarios, TRACE addresses a significant shortcoming in current AV validation methods, ultimately enhancing safety and reliability in autonomous driving technologies.
- TRACE automates accident reconstruction from NHTSA reports into CARLA simulations.
- Utilizes OpenStreetMap for accurate road topology and Large Language Models for vehicle state inference.
- Creates a benchmark of 52 diverse accident scenarios for comprehensive AV testing.
Why It Matters
Improves safety validation for AVs by simulating real-world accident scenarios.