Runtime Input Monitoring of a Reinforcement Learning-Based Agent for Automated Car Parking

Ensuring reliability and safety of the decision-making process of neural architectures for automated driving (AD) is a trendy topic for research. Runtime monitoring algorithms continuously observe a model’s input, detecting anomalies and deviations from the normal data distribution in real-time, so to promptly notify about potential model failure. This paper describes a study to compare a noise-free training environment, in which anomalies are only visible in the test set, and an environment in which out of distribution (OOD) data is present also in validation. We consider two well-known types of noise: Pixel and Gaussian, and three state of the art monitor algorithms: one-class Support Vector Machine (OCSVM), Local Outlier Factor (LOF) and Isolation Forest (IF). As a case study, we take a self-parking vehicle autonomous agent trained through reinforcement learning in the CARLA driving simulator. Results reveal that OCSVM beats LOF and IF in noise-free training, while LOF performs better in the OOD-aware validation case. More research is needed to address generalization, also considering real-world cases, and exploit OOD awareness to increase overall model’s performance.