Student Research Track: Letting the Automated Vehicle Explain Its Own Errors: A Simulator Testbed for Passenger-Facing Explanation HMIs in SOTIF Error Situations
Abstract
A fully automated (SAE L5) vehicle can obey every traffic rule and still behave near-erroneously, and a passenger who cannot tell why loses situation awareness (SA) to the black-box driving stack, degrading the ride. Prior natural-language explanations target drivers or stop at model interpretability, not a passenger-facing account of the vehicle’s own error. This paper designs such an explainable-AI (XAI) interface and builds a simulator testbed to validate it before any human study. We frame a scenario’s error as the functional and output insufficiency (FI/OI) defined by ISO 21448 Safety of the Intended Functionality (SOTIF), and organize a Vision-Language-Action (VLA) model’s reasoning trace into Endsley’s three SA stages. The first contribution is a reusable design that maps this reasoning onto SA stages so the vehicle explains its own FI/OI. The second is a Passenger-in-the-Loop testbed coupling CARLA, a 6-DOF motion platform, and live visual and voice Human-Machine Interfaces (HMIs) over WebSocket, with six metrics as a scenario quality-assurance instrument. In a human-free dry run of 100 replays per scenario, the roundabout-deadlock scenario ran collision-free with deterministic metrics (CV ≤ 0.4%), while the aquaplaning scenario fired its terrain-transition events in every replay and is characterized by its multi-run distribution. We outline a follow-up human-subject study to test affect induction and user experience.
