MuJoCo¶

Ant

Inverted Double Pendulum

MuJoCo stands for Multi-Joint dynamics with Contact. It is a physics engine for facilitating research and development in robotics, biomechanics, graphics and animation, and other areas where fast and accurate simulation is needed. There is physical contact between the robots and their environment - and MuJoCo attempts at getting realistic physics simulations for the possible physical contact dynamics by aiming for physical accuracy and computational efficiency.

The unique dependencies including the MuJoCo simlator for this set of environments can be installed via:

pip install gymnasium[mujoco]

As of October 2021, DeepMind has acquired MuJoCo and has open-sourced it in 2022, making it free for everyone. Using MuJoCo with Gymnasium requires the framework mujoco be installed (this dependency is installed with the above command). Instructions for installing the MuJoCo engine can be found on their website and GitHub repository.

For MuJoCo v3 environments and older the mujoco-py framework is required (pip install gymnasium[mujoco-py]) which can be found in the GitHub repository.

There are eleven MuJoCo environments (in roughly increasing complexity):

Robot	Short Description
CartPoles
InvertedPendulum	MuJoCo version of the CartPole Environment (with Continuous actions)
InvertedDoublePendulum	2 Pole variation of the CartPole Environment
Arms
Reacher	2d arm with the goal of reaching an object
Pusher	3d arm with the goal of pushing an object to a target location
2D Runners
HalfCheetah	2d quadruped with the goal of running
Hopper	2d monoped with the goal of hopping
Walker2d	2d biped with the goal of walking
Swimmers
Swimmer	3d robot with the goal of swimming
Quarduped
Ant	3d quadruped with the goal of running
Humanoid Bipeds
Humanoid	3d humanoid with the goal of running
HumanoidStandup	3d humanoid with the goal of standing up

All of these environments are stochastic in terms of their initial state, with a Gaussian noise added to a fixed initial state in order to add stochasticity. The state spaces for MuJoCo environments in Gymnasium consist of two parts that are flattened and concatenated together: the position of the body part and joints (mujoco.MjData.qpos) and their corresponding velocity (mujoco.MjData.qvel) (more information in the MuJoCo Physics State Documentation).

Among the Gymnasium environments, this set of environments can be considered as more difficult to solve by policy.

Environments can be configured by changing the xml_file argument and/or by tweaking the parameters of their classes.

Versions¶

Gymnasium includes the following versions of the environments:

Version	Simulator	Notes
`v5`	`mujoco=>2.3.3`	Recommended (most features, the least bugs)
`v4`	`mujoco=>2.1.3`	Maintained for reproducibility
`v3`	`mujoco-py`	Deprecated, Kept for reproducibility (limited support)
`v2`	`mujoco-py`	Deprecated, Kept for reproducibility (limited support)

For more information, see the section “Version History” for each environment.

v1 and older are no longer included in Gymnasium.

Comparing training performance across versions¶

The training performance of v2 and v3 is identical assuming the same/default arguments were used.

The training performance of v2/v3 and v4 are not directly comparable because of the change to the newer simulator, but the results for not Ant and not Humanoids are comparable (for more information see GitHub Comment #1 and GitHub Comment #2).

The Training performance of v4 and v5 is different because of the many changes in the environments, but the Half Cheetah and Swimmer exhibits identical behaviour, Pusher and Swimmer are close (for more information see GitHub Issue).

Exact reproducibility¶

Note: The exact behavior of the MuJoCo simulator changes slightly between mujoco versions due to floating point operation ordering (more information of their Documentation), if exact reproducibility is need besides using the seed for experiments the same simulator version should be used.

Rendering Arguments¶

All of the MuJoCo Environments besides the general Gymnasium arguments, and environment specific arguments they also take the following arguments for configuring the renderer:

env = gymnasium.make("Ant-v5", render_mode="rgb_array", width=1280, height=720)

Parameter	Type	Default	Description
`render_mode`	str	`None`	The modality of the render result. Must be one of `human`, `rgb_array`, `depth_array`, or `rgbd_tuple`. Note that `human` does not return a rendered image, but renders directly to the window
`width`	int	`480`	The width of the render window
`height`	int	`480`	The height of the render window
`camera_id`	int \| None	`None`	The camera ID used for the render window
`camera_name`	str \| None	`None`	The name of the camera used for the render window (mutually exclusive option with `camera_id`)
`default_camera_config`	dict[str, float \| int] \| None	`None`	The mjvCamera properties
`max_geom`	int	`1000`	Max number of geometrical objects to render (useful for 3rd-party environments)
`visual_options`	Dict[int, bool]	`{}`	A dictionary with mjVisual flags and value pairs, example `{mujoco.mjtVisFlag.mjVIS_CONTACTPOINT: True, mujoco.mjtVisFlag.mjVIS_CONTACTFORCE: True}` (show contact points and forces).

Rendering Backend¶

The MuJoCo simulator renders images with OpenGL and can use 3 different back ends “glfw” (default), “egl”, “omesa”, which can be selected by setting an environment variable.

Backend	Environment Variable	Description
`GLFW`	`MUJOCO_GL=glfw` (default)	Renders with Window System on GPU
`EGL`	`MUJOCO_GL=egl`	Renders headless on GPU
`OSMESA`	`MUJOCO_GL=osmesa`	Renders headless on CPU

More information of the MuJoCo/OpenGL documentation.