safety_gymnasium

Safety-Gymnasium is a highly modular and easily customizable benchmark environment library based on MuJoCo to facilitate research in the Safe RL domain. Safe Reinforcement Learning (Safe RL) focuses on addressing common issues in reinforcement learning (RL) where agents may exploit reward structures by engaging in reward hacking, risky behaviors, or unrealistic strategies. These behaviors can lead to unintended outcomes, where agents maximize rewards in ways that contradict the original goals of the task. Safe RL aims to ensure that agents not only maximize rewards but also adhere to predefined safety constraints, resulting in more balanced, realistic, and effective policies.

Safety-Gymnasium, built on top of MuJoCo, is a modular and highly customizable benchmark environment specifically designed for Safe RL research. It allows researchers to test and train RL agents in environments where safety is a priority. The framework offers a clean, easy-to-read codebase with well-structured components like agents, tasks, and objects, which makes it simple to modify and create new environments. It also provides support for two new task families, Manipulation and Vision, while including refactored versions of classic Safe RL environments like Safety-Gym and Safety-Velocity.