DeepMind Control Suite

DeepMind's standard set of MuJoCo-based continuous-control reinforcement-learning tasks.

The DeepMind Control Suite was released in 2018 by Yuval Tassa and colleagues at DeepMind, as part of the open-source dm_control library. It uses the MuJoCo physics engine to build a set of continuous-control tasks across domains such as the cart-pole, cheetah, planar biped walker, humanoid, quadruped, and finger, with tasks of varying difficulty within each domain. Every task shares a uniform structure and an interpretable reward: each step's reward falls between 0 and 1, and every episode runs a fixed 1,000 steps, so a perfect score is 1,000. Input can be either joint states or raw camera pixels alone, which is why it became a standard benchmark for sample efficiency and for “learning control from pixels,” with algorithms such as Dreamer, DrQ, and TD-MPC all reporting results on it. dm_control also ships with tools for editing MJCF models, building environments from composable parts, and multi-agent soccer.

ExampleIn the walker-walk task, a planar biped model learns to walk forward using only camera images as input, and different algorithms are compared on how much return (up to a ceiling of 1,000) they achieve within the same number of interaction steps.

Also called

DMC, dm_control, DM Control

Related

Reinforcement Learning · MuJoCo (Multi-Joint dynamics with Contact) · Benchmark · DreamerV3 · TD-MPC2 · Sample Efficiency

Sources

DeepMind Control Suite (arXiv 1801.00690)
google-deepmind/dm_control GitHub