Swimmer#

This environment is part of the Mujoco environments. Please read that page first for general information.


Action Space	Box(-1.0, 1.0, (2,), float32)
Observation Shape	(8,)
Observation High	[inf inf inf inf inf inf inf inf]
Observation Low	[-inf -inf -inf -inf -inf -inf -inf -inf]
Import	`gym.make("Swimmer-v4")`

Description#

This environment corresponds to the Swimmer environment described in Rémi Coulom’s PhD thesis “Reinforcement Learning Using Neural Networks, with Applications to Motor Control”. The environment aims to increase the number of independent state and control variables as compared to the classic control environments. The swimmers consist of three or more segments (’links’) and one less articulation joints (’rotors’) - one rotor joint connecting exactly two links to form a linear chain. The swimmer is suspended in a two dimensional pool and always starts in the same position (subject to some deviation drawn from an uniform distribution), and the goal is to move as fast as possible towards the right by applying torque on the rotors and using the fluids friction.

Notes#

The problem parameters are: Problem parameters:

n: number of body parts
m_i: mass of part i (i ∈ {1…n})
l_i: length of part i (i ∈ {1…n})
k: viscous-friction coefficient

While the default environment has n = 3, l_i = 0.1, and k = 0.1. It is possible to pass a custom MuJoCo XML file during construction to increase the number of links, or to tweak any of the parameters.

Action Space#

The action space is a Box(-1, 1, (2,), float32). An action represents the torques applied between links

Num	Action	Control Min	Control Max	Name (in corresponding XML file)	Joint	Unit
0	Torque applied on the first rotor	-1	1	rot2	hinge	torque (N m)
1	Torque applied on the second rotor	-1	1	rot3	hinge	torque (N m)

Observation Space#

By default, observations consists of:

θ_i: angle of part i with respect to the x axis
θ_i’: its derivative with respect to time (angular velocity)

In the default case, observations do not include the x- and y-coordinates of the front tip. These may be included by passing exclude_current_positions_from_observation=False during construction. Then, the observation space will have 10 dimensions where the first two dimensions represent the x- and y-coordinates of the front tip. Regardless of whether exclude_current_positions_from_observation was set to true or false, the x- and y-coordinates will be returned in info with keys "x_position" and "y_position", respectively.

By default, the observation is a ndarray with shape (8,) where the elements correspond to the following:

Num	Observation	Min	Max	Name (in corresponding XML file)	Joint	Unit
0	angle of the front tip	-Inf	Inf	rot	hinge	angle (rad)
1	angle of the first rotor	-Inf	Inf	rot2	hinge	angle (rad)
2	angle of the second rotor	-Inf	Inf	rot3	hinge	angle (rad)
3	velocity of the tip along the x-axis	-Inf	Inf	slider1	slide	velocity (m/s)
4	velocity of the tip along the y-axis	-Inf	Inf	slider2	slide	velocity (m/s)
5	angular velocity of front tip	-Inf	Inf	rot	hinge	angular velocity (rad/s)
6	angular velocity of first rotor	-Inf	Inf	rot2	hinge	angular velocity (rad/s)
7	angular velocity of second rotor	-Inf	Inf	rot3	hinge	angular velocity (rad/s)

Rewards#

The reward consists of two parts:

forward_reward: A reward of moving forward which is measured as forward_reward_weight * (x-coordinate before action - x-coordinate after action)/dt. dt is the time between actions and is dependent on the frame_skip parameter (default is 4), where the frametime is 0.01 - making the default dt = 4 * 0.01 = 0.04. This reward would be positive if the swimmer swims right as desired.
ctrl_cost: A cost for penalising the swimmer if it takes actions that are too large. It is measured as ctrl_cost_weight * sum(action²) where ctrl_cost_weight is a parameter set for the control and has a default value of 1e-4

The total reward returned is reward = forward_reward - ctrl_cost and info will also contain the individual reward terms

Starting State#

All observations start in state (0,0,0,0,0,0,0,0) with a Uniform noise in the range of [-reset_noise_scale, reset_noise_scale] is added to the initial state for stochasticity.

Episode End#

The episode truncates when the episode length is greater than 1000.

Arguments#

No additional arguments are currently supported in v2 and lower.

gym.make('Swimmer-v4')

v3 and v4 take gym.make kwargs such as xml_file, ctrl_cost_weight, reset_noise_scale etc.

env = gym.make('Swimmer-v4', ctrl_cost_weight=0.1, ....)

Parameter	Type	Default	Description
`xml_file`	str	`"swimmer.xml"`	Path to a MuJoCo model
`forward_reward_weight`	float	`1.0`	Weight for forward_reward term (see section on reward)
`ctrl_cost_weight`	float	`1e-4`	Weight for ctrl_cost term (see section on reward)
`reset_noise_scale`	float	`0.1`	Scale of random perturbations of initial position and velocity (see section on Starting State)
`exclude_current_positions_from_observation`	bool	`True`	Whether or not to omit the x- and y-coordinates from observations. Excluding the position can serve as an inductive bias to induce position-agnostic behavior in policies

Version History#

v4: all mujoco environments now use the mujoco bindings in mujoco>=2.1.3
v3: support for gym.make kwargs such as xml_file, ctrl_cost_weight, reset_noise_scale etc. rgb rendering comes from tracking camera (so agent does not run away from screen)
v2: All continuous control environments now use mujoco_py >= 1.50
v1: max_time_steps raised to 1000 for robot based tasks. Added reward_threshold to environments.
v0: Initial versions release (1.0.0)