Swimmer

This Environment is part of MaMuJoCo environments. Please read that page first for general information. The task is Gymansium’s MuJoCo/Swimmer.

Action Space

The shape of the action space depends on the partitioning. The partitioning has the following form:

if partitioning is None:

Instantiate	env = mamujoco_v1.parallel_env("Swimmer", None)
Agents	agents= ['agent_0']
Number of Agents	1
Action Spaces	{'agent_0' : Box(-1, 1, (2,), float32)}
Part partition	(joint0, joint1,)

If partitioning, is None then the environment contains a single agent with the same action space as Gymansium’s MuJoCo/Swimmer.

Num	Action	Control Min	Control Max	Name (in corresponding XML file)	Joint	Unit
0	Torque applied on the first rotor	-1	1	motor1_rot	hinge	torque (N m)
1	Torque applied on the second rotor	-1	1	motor2_rot	hinge	torque (N m)

if partitioning == “2x1”: # isolate upper and lower body

Instantiate	env = mamujoco_v1.parallel_env("Swimmer", "2x1")
Agents	agents= ['agent_0', 'agent_1']
Number of Agents	2
Action Spaces	{'agent_0': Box(-1, 1, (1,), float32), 'agent_1' : Box(-1, 1, (1,), float32)}
Part partition	[(joint0,), (joint1,)]

The environment is partitioned in 2 parts, one part corresponding to the first joint, and one part corresponding to the second joint.

Agent 0 action space

Num	Action	Control Min	Control Max	Name (in corresponding XML file)	Joint	Unit
0	Torque applied on the first rotor	-1	1	motor1_rot	hinge	torque (N m)

Agent 1 action space

Num	Action	Control Min	Control Max	Name (in corresponding XML file)	Joint	Unit
0	Torque applied on the second rotor	-1	1	motor2_rot	hinge	torque (N m)

Observation Space

Observation Categories
Default local_categories	[["qpos", "qvel"], ["qpos"]]
Default global_categories	("qpos", "qvel")
Supported observation categories	"qpos", "qvel"

Besides the local observation of each agent (which depend on their parts of the agent, the observation categories and the observation depth), each agent also observes the position and velocity items of the swimmer’s tip. See more at the Gymnasium’s Swimmer.

Rewards

All agents receive the same Gymnasium’s Swimmer reward.

Starting state

The starting state of the environment is the same as Gymnasium’s Swimmer.

Episode End

All agent terminate and truncate at the same time, given the same conditions as Gymnasium’s Swimmer.

Version History

• v1:

  • Now based on Gymnasium/MuJoCo-v5 instead of Gymnasium/MuJoCo-v4 (https://github.com/Farama-Foundation/Gymnasium/pull/572).

• v0: Initial version release, uses Gymnasium.MuJoCo-v4, and is a fork of the original MaMuJoCo schroederdewitt/multiagent_mujoco. Changes from the original MaMuJoCo (schroederdewitt/multiagent_mujoco):

  • Added/Fixed Global observations (The Swimmer’s front tip: free_body_rot) not being observed.