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 tweak the MuJoCo XML files to increase the number of links, or to tweak any of the parameters.

Action Space

The agent take a 2-element vector for actions. The action space is a continuous (action, action) in [-1, 1], where action represents the numerical 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

The state space consists of:

• A₀: position of first point

• θ_i: angle of part i with respect to the x axis

• A₀, θ_i: their derivatives with respect to time (velocity and angular velocity)

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	x-coordinate of the front tip	-Inf	Inf	slider1	slide	position (m)
1	y-coordinate of the front tip	-Inf	Inf	slider2	slide	position (m)
2	angle of the front tip	-Inf	Inf	rot	hinge	angle (rad)
3	angle of the second rotor	-Inf	Inf	rot2	hinge	angle (rad)
4	angle of the second rotor	-Inf	Inf	rot3	hinge	angle (rad)
5	velocity of the tip along the x-axis	-Inf	Inf	slider1	slide	velocity (m/s)
6	velocity of the tip along the y-axis	-Inf	Inf	slider2	slide	velocity (m/s)
7	angular velocity of front tip	-Inf	Inf	rot	hinge	angular velocity (rad/s)
8	angular velocity of second rotor	-Inf	Inf	rot2	hinge	angular velocity (rad/s)
9	angular velocity of third rotor	-Inf	Inf	rot3	hinge	angular velocity (rad/s)

Note: In practice (and Gym implementation), the first two positional elements are omitted from the state space since the reward function is calculated based on those values. Therefore, observation space has shape (8,) and looks like:

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 second 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 second rotor	-Inf	Inf	rot2	hinge	angular velocity (rad/s)
7	angular velocity of third rotor	-Inf	Inf	rot3	hinge	angular velocity (rad/s)

Rewards

The reward consists of two parts:

• reward_front: A reward of moving forward which is measured as (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 dt for one frame is 0.01 - making the default dt = 4 * 0.01 = 0.04. This reward would be positive if the swimmer swims right as desired.

• reward_control: A negative reward for penalising the swimmer if it takes actions that are too large. It is measured as -coefficient x sum(action²) where coefficient is a parameter set for the control and has a default value of 0.0001

The total reward returned is reward = reward_front + reward_control

Starting State

All observations start in state (0,0,0,0,0,0,0,0) with a Uniform noise in the range of [-0.1, 0.1] is added to the initial state for stochasticity.

Episode Termination

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

Arguments

No additional arguments are currently supported (in v2 and lower), but modifications can be made to the XML file in the assets folder (or by changing the path to a modified XML file in another folder).

gym.make('Swimmer-v2')

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, ....)

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)