Adroit

A benchmark for controlling a 24-DoF simulated five-fingered hand in MuJoCo to open doors, hammer nails, and more, across four tasks.

The Adroit tasks come from a 2018 paper by Aravind Rajeswaran, Vikash Kumar, Sergey Levine, and colleagues (University of Washington, OpenAI, Berkeley), simulating a 24-degree-of-freedom ADROIT anthropomorphic hand in MuJoCo (D4RL refers to it as a simulated Shadow Hand) across four tasks: moving a ball to a target location, spinning a pen in-hand to a target orientation, hammering a nail, and pulling open a latched door. The authors collected 25 human demonstrations per task using a VR data glove, and proposed DAPG: behavioral cloning on the demonstrations first, followed by reinforcement-learning fine-tuning with a policy-gradient method that incorporates the demonstrations. D4RL later incorporated these tasks into its offline reinforcement-learning benchmark, providing human, cloned, and expert data for each, and Gymnasium-Robotics also maintains these environments. They are commonly used to test high-dimensional dexterous manipulation, sparse rewards, and demonstration-assisted learning.

ExampleD4RL's pen-human dataset consists of the 25 human demonstrations for the pen-spinning task, and offline reinforcement-learning papers commonly use it to compare algorithms.

Also called

Adroit Hand, ADROIT, DAPG Tasks

Related

Dexterous Manipulation · In-hand Manipulation · Shadow Dexterous Hand · D4RL · MuJoCo (Multi-Joint dynamics with Contact) · Sparse Reward

Sources

Learning Complex Dexterous Manipulation with Deep Reinforcement Learning and Demonstrations (arXiv 1709.10087)
D4RL: Datasets for Deep Data-Driven Reinforcement Learning (arXiv 2004.07219)
Gymnasium-Robotics: Adroit Hand