Farama-Foundation/Minari

What it solves

Minari provides a standardized way to manage, share, and load datasets for offline reinforcement learning (RL). It acts as a central hub for RL data, similar to how HuggingFace's datasets library works for NLP, allowing researchers to easily access pre-recorded experience data without needing to interact with a live environment.

How it works

Minari allows users to download remote datasets via a command-line interface or load them directly into Python using minari.load_dataset(). It also provides a DataCollector wrapper for Gymnasium environments, which enables users to record their own agent interactions (observations, actions, rewards, and terminations) and save them as new datasets.

Who it’s for

It is designed for researchers and developers working on offline reinforcement learning who need a consistent format for handling experience datasets.

Highlights

• Standardized Data Loading: Simple API for iterating through episodes and accessing RL-specific data like rewards and observations.
• Dataset Creation: Built-in tools to record data from Gymnasium environments to create new offline datasets.
• CLI Tools: Command-line utilities to list, show, and { "off_topic": false, "markdown": "## What it solves Minari provides a standardized way to manage, share, and load datasets for offline reinforcement learning (RL). It acts as a central hub for RL data, similar to how HuggingFace's datasets library works for NLP, allowing researchers to easily access pre-recorded experience data without needing to interact with a live environment.

How it works

Minari allows users to download remote datasets via a command-line interface or load them directly into Python using minari.load_dataset(). It also provides a DataCollector wrapper for Gymnasium environments, which enables users to record their own agent interactions (observations, actions, rewards, and terminations) and save them as new datasets.

Who it’s for

It is designed for researchers and developers working on offline reinforcement learning who need a consistent format for handling experience datasets.

Highlights

• Standardized Data Loading: Simple API for iterating through episodes and accessing RL-specific data like rewards and observations.
• Dataset creation: Built-in tools to record data from Gymnasium environments to create new offline datasets.
• CLI Tools: Command-line utilities to list, show, and download datasets.