Reinforcement learning (RL) has been successfully applied to sequential decision-making problems, e.g., playing computer games or solving robotic tasks in simulations. However, RL methods are not yet ready to be applied to real robotic systems if safety is a major concern. To address this issue, we propose a safety layer based on control barrier functions to ensure safety for an RL-based motion planner for highway scenarios with a continuous action space. Our method ensures legal safety by following traffic rules. Moreover, we propose a relaxation mechanism so that safety is restored as soon as possible when other vehicles violate traffic rules and render our optimization problem infeasible. We evaluate our approach on a real-world highway dataset and a traffic simulator. Numerical experiments confirm that an agent equipped with our proposed safety layer does not cause any accidents during learning and yet reaches the goal as often as an agent without a safety layer.     «

Reinforcement learning (RL) has been successfully applied to sequential decision-making problems, e.g., playing computer games or solving robotic tasks in simulations. However, RL methods are not yet ready to be applied to real robotic systems if safety is a major concern. To address this issue, we propose a safety layer based on control barrier functions to ensure safety for an RL-based motion planner for highway scenarios with a continuous action space. Our method ensures legal safety by follo...     »