Human-in-the-loop Learning for Safe Exploration through Anomaly Prediction and Intervention

Deep-learning based approaches for learning autonomous driving policies comes with a set of safety challenges. Human-in-the-loop (HITL) learning can be used to improve the safety and reliability of such systems by embedding the human understanding of the complex notion of safety. As AI systems are increasingly deployed in situations with real-world consequences for humans, it can be beneficial to involve humans in various stages of the life-cycle of AI systems to ensure safe and compliant behavior by the systems. In this position paper, we propose a new method to incorporate human-in-the-loop learning to facilitate safe exploration.