This research paper presents Safety-Gymnasium, an environment suite for safety-critical tasks in both single and multi-agent scenarios in AI learning. It also introduces a library called Safe Policy Optimization, which includes 16 state-of-the-art SafeRL algorithms. The aim is to facilitate the evaluation and comparison of safety performance in AI systems, promoting the development of reinforcement learning for safer and more reliable real-world applications. The paper emphasizes the necessity of SafeRL in high-stake domains like autonomous vehicles and healthcare, where system failures can have severe consequences.
Publication date: 19 Oct 2023
Project Page: https://sites.google.com/view/safety-gymnasium
Paper: https://arxiv.org/pdf/2310.12567
