PKU-Alignment/safety-gymnasium

What it solves

Safety-Gymnasium provides a standardized, scalable, and customizable benchmark for Safe Reinforcement Learning (SafeRL). It addresses the need for a unified set of environments that can be used to evaluate SafeRL algorithms, replacing older libraries that lacked support for vectorized environments, modern Gym APIs, or vision-based inputs.

How it works

Built on top of MuJoCo (2.3.0+), the library removes the dependency on the legacy mujoco-py and implements a wide array of safety-enhanced learning tasks. It provides standard APIs that include information on constraints, allowing agents to learn tasks while adhering to safety requirements. The library supports various modalities, including state-based and vision-based inputs, and includes environments for single-agent, multi-agent, and Isaac Gym simulations.

Who it’s for

It is designed for researchers and developers working on Safe Reinforcement Learning who need a reliable, modern benchmark to test and compare the performance and safety of their algorithms.

Highlights

• Broad Environment Suite: Includes tasks for safe navigation (Goal, Button, Push, Circle), safe velocity, safe vision, safe multi-agent, and Safe Isaac Gym.
• Modern Infrastructure: Supports vectorized environments and is compatible with the new Gym API (0.26.0+).
• Diverse Agents: Features multiple agent types including Point, Car, Doggo, Racecar, Ant, and others like HalfCheetah and Humanoid.
• Improved Stability: Redesigns the original Safety-Gym to fix bugs and remove outdated dependencies.