Gym/Gymnasium MuJoCo Tasks

A classic set of MuJoCo-based continuous-control environments in Gymnasium, a standard test bed for reinforcement-learning papers.

This is a set of 11 MuJoCo physics-simulation environments, originally released with OpenAI Gym and now maintained by the Farama Foundation's Gymnasium; the most commonly used are HalfCheetah (a 2D running cheetah), Hopper (a single-leg hopper), Walker2d (a 2D biped walker), Ant (a quadruped), and Humanoid. Actions are continuous joint torques, the reward is roughly “move forward as fast as possible, minus energy spent on actions,” and each episode runs at most 1,000 steps. Reinforcement-learning algorithms such as SAC, TD3, and PPO all report scores on them, and several of the D4RL offline datasets are also built on top of them. They test only locomotion control, with no vision or real-robot detail involved, making them well suited for learning the ropes and comparing algorithms. Version v5 is currently recommended (requires mujoco≥2.3.3), and scores from different versions cannot be compared directly.

ExampleHalfCheetah-v5: the action is 6 joint torques (a 6-dimensional vector), the observation is 17-dimensional joint positions and velocities, there's no fall-based termination, the episode truncates at 1,000 steps, and the score equals a forward-velocity reward minus a control cost.

Also called

Gymnasium MuJoCo Environments, MuJoCo Locomotion Tasks, HalfCheetah / Hopper / Walker2d

Related

MuJoCo (Multi-Joint dynamics with Contact) · Gymnasium · D4RL · Soft Actor-Critic · Proximal Policy Optimization · DeepMind Control Suite

Sources

Gymnasium Documentation: MuJoCo environments
Gymnasium Documentation: Half Cheetah

As of

2026-09