Door#

These datasets were generated with the AdroitHandDoor-v1 environment, originally hosted in the hand_dapg repository. The objective of the task is to open a door with a 24-DoF robotic hand. This domain was selected to measure the effect of a narrow expert data distributions and human demonstrations on a sparse reward, high-dimensional robotic manipulation task.

There are three types of datasets, two from the original paper[1] (human and expert), and another one introduced in D4RL[2] (cloned).

References#

[1] Rajeswaran, Aravind, et al. ‘Learning Complex Dexterous Manipulation with Deep Reinforcement Learning and Demonstrations’. CoRR, vol. abs/1709.10087, 2017, http://arxiv.org/abs/1709.10087.

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

Available Datasets#

Dataset ID

Description

door-cloned-v1

Data obtained by training an imitation policy on the demonstrations from expert and human, then running the policy, and mixing data at a 50-50 ratio with the demonstrations

door-expert-v1

Trajectories have expert data from a fine-tuned RL policy provided in the DAPG repository

door-human-v1

25 human demonstrations provided in the DAPG repository