search

Farama-Foundation / minari-dataset-generation-scripts

Scripts to recreate the D4RL datasets with Minari
MIT License
18 stars 5 forks source link

readme

D4RL to Minari datasets

This repository contains the scripts to convert the D4RL environment datasets based on MuJoCo to Minari datasets.

The environments used to regenerate the datasets are refactored versions of the originals. These new environment versions are now maintained in the Gymnasium-Robotics project, they follow the Gymnasium API, and have been updated to use the latest mujoco bindings from Deepmind.

Installation

git clone https://github.com/Farama-Foundation/minari-dataset-generation-scripts

For reproducibility, each set of dataset creation scripts has its own requirements. For example, the Point Maze requirements can be installed with pip install -r scripts/pointmaze/requirements.txt.

Create datasets

Point Maze

The point maze datasets have been regenerated using the same q_iteration expert policy as in the original D4RL paper. The environments used can be found here.

You can run the script used to regenerate the datasets with:

python scripts/pointmaze/create_pointmaze_dataset --env "PointMaze_UMaze-v3" --dataset_name="pointmaze-umaze-v0" --maze-solver="QIteration"

This will generate a local Minari dataset named pointmaze-umaze-v0 for the PointMaze_UMaze-v3 environment, using q_iteration as the expert policy, Depth First Search can also be used as the algorithm to generate a path to the goal by passing "DFS" instead of "QIteration".

Adroit Hand

The Minari datasets for the Adroit Hand environments are recreated by reading the original D4RL datasets and adapting them to the Minari standards. The dataset versions from D4RL are v1 and the new Minari datasets name are relabeled to version v0. The datasets for the environments include human, expert, and cloned, and they can be created as follows:

AdroitHandDoor-v1:

python scripts/adroit/recreate_adroit_door.py

AdroitHandHammer-v1:

python scripts/adroit/recreate_adroit_hammer.py

AdroitHandPen-v1:

python scripts/adroit/recreate_adroit_pen.py

AdroitHandRelocate-v1:

python scripts/adroit/recreate_adroit_relocate.py

Ant Maze

To create the dataset, a goal reaching expert policy was previously trained with the SAC algorithm provided in Stable Baselines 3. This goal reaching policy is then used by the Ant agent to follow a set of waypoints generated by a planner (QIteration) to the final goal location. Six different datasets are available, corresponding to the different configurations of maps and goal/reset positions, as in the original D4RL paper. These can be recreated by changing the working directory to scripts/antmaze and then running

python create_antmaze_dataset.py

Minigrid

This dataset generated from the MiniGrid-FourRooms environment. The objective of the agent is to reach a goal position in a gridworld. We regenerate the dataset of D4RL for full reproducibility, using a random policy and an expert policy that navigates straight to the goal. The datasets can be recreated as follows:

minigrid-fourrooms-v0:

python scripts/minigrid/generate_minigrid.py

minigrid-fourrooms-random-v0:

python scripts/minigrid/generate_minigrid.py --random

FrankaKitchen

The goal of the FrankaKitchen environment is to interact with the various objects in order to reach a desired state configuration. These datasets were generated with the FrankaKitchen-v1 environment, originally hosted in the D4RL and relay-policy-learning repository. These datasets can be recreated using:

python scripts/kitchen/create_franka_dataset.py

More datasets to come

Checking datasets

This repository also provides scripts that run simple checks on the integrity of Minari datasets. For example, generic checks on a dataset can be run using python checks/check_dataset.py <dataset_id>.