Farama-Foundation/Gymnasium
Gymnasium
Gymnasium is a suite of standardized APIs and simulation toolkits used to evaluate agent behavior and benchmark reinforcement learning algorithms. It provides a standardized interface for creating and interacting with simulated environments, enabling the training of reinforcement learning agents through a consistent set of interaction protocols.
The project emphasizes experimental reproducibility through a versioned API and a system for tracking changes to environment logic using version suffixes. This ensures that learning results remain consistent and can be replicated across different software releases.
The toolkit includes a collection of reference simulation tasks across physics, text, and game-based scenarios. It supports vectorized environment execution to collect experience in parallel, as well as wrapper-based logic to modify observations or rewards without altering the primary simulation engine.
Features
- Agent Simulation Environments - Provides a toolkit of physics, text, and game-based scenarios to test and evaluate agent behaviors.
- State-Action-Reward Interfaces - Implements a synchronous cycle where agents provide actions and environments return observations and rewards.
- Autonomous Agent Simulations - Allows agents to be run through diverse physics, text, and game-based scenarios to evaluate control complexity.
- Performance Benchmarking - Provides tools for evaluating and comparing agent performance across a standardized set of simulation environments.
- Reinforcement Learning Environments - Provides a standardized interface for defining state, action, and reward logic to train autonomous agents.
- RL Reproducibility Standards - Ensures that RL training trials yield consistent results across different software releases via simulation logic tracking.
- RL Interface Standards - Provides universal communication protocols and API specifications for interactions between reinforcement learning agents and environments.
- RL Reference Environments - Provides a collection of established simulation environments following a common API for benchmarking learning algorithms.
- Vectorized Environments - Provides execution wrappers that run multiple simulation instances in parallel to accelerate experience collection.
- Environment Version Locking - Uses version suffixes in environment names to pin simulation logic and ensure result reproducibility.
- Observation and Reward Wrappers - Includes wrapper-based logic to modify observations or rewards without altering the primary simulation engine.
- RL Simulation Tasks - Ships a comprehensive library of diverse simulation tasks that adhere to a common communication protocol.
- Machine Learning - Standardized environments for reinforcement learning algorithms.
- Reinforcement Learning - Standard API for single-agent reinforcement learning environments.
