siddargu / minari

A standard format for offline reinforcement learning datasets, with popular reference datasets and related utilities

Minari is the new name of this library. Minari used to be called Kabuki.

Minari is intended to be a Python library for conducting research in offline reinforcement learning, akin to an offline version of Gymnasium or an offline RL version of HuggingFace's datasets library. This library is currently in beta, and we're targeting having a complete initial release in February

We have a public discord server (which we also use to coordinate development work) that you can join here: https://discord.gg/jfERDCSw.

Installation

pip install numpy cython

pip install git+https://github.com/Farama-Foundation/Minari.git

Downloading datasets

import minari

dataset = minari.download_dataset("LunarLander_v2_remote-test-dataset")

Recreating Gymnasium environments (Coming very soon!)

import json
import gymnasium as gym
from gymnasium.utils.serialize_spec_stack import deserialise_spec_stack

env = gym.make(deserialise_spec_stack(json.loads(dataset.environment_stack)))

Uploading datasets

dataset.save()
dataset = minari.upload_dataset("LunarLander_v2_remote-test-dataset")

Saving to dataset format

It is not the aim of Minari to insist that you use a certain buffer implementation. However, in order to maintain standardisation across the library, we have a standardised format, the MinariDataset class, for saving replay buffers to file.

This converter will have tests to ensure formatting standards

Checking available remote datasets

import minari

minari.list_remote_datasets()

Checking available local datasets

import minari
minari.list_local_datasets()

---

Minari is a shortening of Minarai, the Japanese word for "learning by observation".