D4RL¶

The D4RL dataset group contains a reproduction of the datasets from the D4RL benchmark[1]. For reproducibility purposes, not all the datasets are the same as in D4RL, but they are generated with the same principles. We provide the code that reproduces each dataset on GitHub in the repository Farama-Foundation/minari-dataset-generation-scripts.

References¶

[1] Fu, Justin, et al. ‘D4RL: Datasets for Deep Data-Driven Reinforcement Learning’. CoRR, vol. abs/2004.07219, 2020, https://arxiv.org/abs/2004.07219.

Content¶

ID	Description
Ant Maze	The Ant Maze datasets present a navigation domain that replaces the 2D ball from pointmaze with the more complex 8-DoF Ant quadruped robot
Door	These datasets were generated with the `AdroitHandDoor-v1` environment, originally hosted in the `hand_dapg` repository
Hammer	These datasets were generated with the `AdroitHandHammer-v1` environment, originally hosted in the `hand_dapg` repository
Kitchen	These datasets were generated with the `FrankaKitchen-v1` environment, originally hosted in the `D4RL`[1] and `relay-policy-learning`[2] repository
MiniGrid	Dataset generated from the MiniGrid-FourRooms environment
Pen	These datasets were generated with the `AdroitHandPen-v1` environment, originally hosted in the `hand_dapg` repository
Point Maze	The Point Maze domain involves moving a force-actuated ball (along the X and Y axis) to a fixed target location
Relocate	These datasets were generated with the `AdroitHandRelocate-v1` environment, originally hosted in the `hand_dapg` repository