Abstract

The DeepMind Control Suite is a set of continuous control tasks with astandardised structure and interpretable rewards, intended to serve asperformance benchmarks for reinforcement learning agents. The tasks are writtenin Python and powered by the MuJoCo physics engine, making them easy to use andmodify. We include benchmarks for several learning algorithms. The ControlSuite is publicly available at https://www.github.com/deepmind/dm_control. Avideo summary of all tasks is available at http://youtu.be/rAai4QzcYbs .

 

Quick Read (beta)