Transformer-based out-of-distribution (OOD) input detection for trustworthy self-parking

Abstract

Abstract The deployment of automated driving (AD) systems demands trustworthiness, explainability, and accountability. A challenge is detecting out-of-distribution (OOD) inputs from sensor noise, hardware failures, adversarial interference, or deviations from the operational design domain. Common approaches use OOD samples during training or threshold selection, potentially limiting generalization beyond the anomaly types represented during development. We argue that OOD-unaware novelty detection is crucial for detection under unforeseen conditions. This paper enhances a deep-reinforcement-learning-based self-parking agent developed in CARLA by integrating a novel OOD detector. We present Time-Frequency-Memory-enhanced (TFMe), a dual-branch, memory-augmented, encoder-only Transformer that outperforms the evaluated baselines. In OOD-unaware settings, decision thresholds are calibrated exclusively on in-distribution validation data using a 99th-percentile rule. We assess sensitivity across multiple noise, attack types, and severity levels. Results show that OOD-unaware calibration maintains consistent performance across anomaly types and intensities, avoiding the drops observed under OOD-aware calibration. Experiments on real-world data from Lyft with synthetically injected anomalies further show that OOD-unaware calibration transfers more effectively to the evaluated unseen anomaly types. Transferability across parking environments is assessed on an unseen 60° angled layout without retraining. We show that the TFMe-enhanced ADF mitigates collision risk and reduces timeouts in the evaluated parking scenarios by issuing Take-Over Requests after repeated anomaly detections. TFMe has lower inference latency than the self-parking agent, allowing it to operate in parallel without extending the estimated critical path. Finally, compared with Raspberry Pi 5, deployment on a Jetson Orin Nano achieves 19% lower inference time and 30% lower energy consumption.