This article explores the use of shielding in reinforcement learning (RL) for continuous environments. Classical shielding approaches have limitations, making them hard to use in complex environments. The authors extend the approximate model-based shielding (AMBS) framework to these continuous settings, using Safety Gym as the test-bed. They offer strong probabilistic safety guarantees and propose two new penalty techniques that modify the policy gradient. The results show stable convergence. The paper is a significant contribution to the field of RL, offering ways to ensure safety in complex, continuous environments.
Publication date: 2 Feb 2024
Project Page: N/A
Paper: https://arxiv.org/pdf/2402.00816
