This repository provides Safety Gym, a suite of tools and environments for accelerating research in safe exploration for reinforcement learning. It is designed for researchers and practitioners in RL who need to evaluate and develop algorithms that adhere to safety constraints during learning and execution.

How It Works

Safety Gym leverages the MuJoCo physics simulator to create a variety of robotic manipulation and navigation tasks. It introduces the concept of "constraints" which are quantifiable safety measures that agents must avoid violating. The core innovation lies in its flexible Engine class, allowing users to construct custom environments by specifying robot types, tasks, sensor configurations (including lidar and vision), object placements, and explicit safety constraints. This modular design facilitates systematic benchmarking of safe RL algorithms.

Quick Start & Requirements

• Installation: Clone the repository and install using pip install -e . after installing mujoco_py.
• Prerequisites: Python 3.6+, mujoco_py. Tested on macOS Mojave and Ubuntu 16.04 LTS.
• Resources: Requires MuJoCo installation, which may involve obtaining a license from the MuJoCo website.
• Documentation: https://github.com/openai/safety-gym

Highlighted Details

• Offers a benchmark suite with predefined robot (Point, Car, Doggo), task (Goal, Button, Push), and difficulty levels.
• Supports custom environment creation with configurable robots, tasks, sensors (lidar, vision), and object interactions.
• Provides detailed control over object placement, keep-out zones, and constraint enforcement for fine-grained safety specification.
• Includes mechanisms for comparing algorithms using normalized benchmark scores based on reference statistics.

Maintenance & Community

The project is marked as "Archive" and no further updates are expected. It was developed by OpenAI.

Licensing & Compatibility

The repository does not explicitly state a license in the README. However, OpenAI's typical practice for research code is to release under a permissive license like MIT. Users should verify the license for commercial use.

Limitations & Caveats

The project is archived, meaning no further development or bug fixes are anticipated. Vision support is minimally implemented and considered low-priority. Random layout generation can occasionally fail to produce a valid scene configuration.