stable_gym.envs.mujoco.half_cheetah_cost

Modified version of the HalfCheetah Mujoco environment found in the :gymnasiumL`gymnasium library <environments/mujoco/half_cheetah>`. This modification was first described by Han et al. 2020. In this modified version:

• The objective was changed to a velocity-tracking task. To do this, the reward is replaced with a cost. This cost is the squared difference between the HalfCheetah’s forward velocity and a reference value (error).

• Three optional variables were added to the observation space; The reference velocity, the reference error (i.e. the difference between the cheetah’s forward velocity and the reference) and the cheetah’s forward velocity. These variables can be enabled using the exclude_reference_from_observation, exclude_reference_error_from_observation and exclude_velocity_from_observation environment arguments.

Submodules

• stable_gym.envs.mujoco.half_cheetah_cost.half_cheetah_cost

Classes

HalfCheetahCost

Custom HalfCheetah gymnasium environment.

Package Contents

class stable_gym.envs.mujoco.half_cheetah_cost.HalfCheetahCost(reference_forward_velocity=1.0, randomise_reference_forward_velocity=False, randomise_reference_forward_velocity_range=(0.5, 1.5), forward_velocity_weight=1.0, include_ctrl_cost=False, ctrl_cost_weight=0.0001, reset_noise_scale=0.1, exclude_current_positions_from_observation=True, exclude_reference_from_observation=False, exclude_reference_error_from_observation=True, exclude_x_velocity_from_observation=False, action_space_dtype=np.float32, observation_space_dtype=np.float64, **kwargs)

Bases: gymnasium.envs.mujoco.half_cheetah_v4.HalfCheetahEnv, gymnasium.utils.EzPickle

Custom HalfCheetah gymnasium environment.

Note

Can also be used in a vectorized manner. See the gym.vector documentation.

Source:

This is a modified version of the HalfCheetah Mujoco environment found in the gymnasium library. This modification was first described by Han et al. 2020. Compared to the original HalfCheetah environment in this modified version:

• The objective was changed to a velocity-tracking task. To do this, the reward is replaced with a cost. This cost is the squared difference between the HalfCheetah’s forward velocity and a reference value (error). Additionally, also a control cost can be included in the cost.

• Three optional variables were added to the observation space; The reference velocity, the reference error (i.e. the difference between the cheetah’s forward velocity and the reference) and the cheetah’s forward velocity. These variables can be enabled using the exclude_reference_from_observation, exclude_reference_error_from_observation and exclude_velocity_from_observation environment arguments.

The rest of the environment is the same as the original HalfCheetah environment. Below, the modified cost is described. For more information about the environment (e.g. observation space, action space, episode termination, etc.), please refer to the gymnasium library.

Important

The original code from Han et al. 2020 terminates the episode if the cheetah’s back thigh angle exceeds \(0.5 \pi\) or falls below \(-0.5 \pi\). This condition, not mentioned in the paper, is not implemented here as it’s not part of the original environment.

Modified cost:

A cost, computed using the HalfCheetahCost.cost() method, is given for each simulation step, including the terminal step. This cost is defined as the error between the Cheetah’s forward velocity and a reference value. A control cost can also be included in the cost. The cost is computed as:

\[cost = w_{forward\_velocity} \times (x_{velocity} - x_{reference\_x\_velocity})^2 + w_{ctrl} \times c_{ctrl}\]

Solved Requirements:

Considered solved when the average cost is less than or equal to 50 over 100 consecutive trials.

How to use:

import stable_gym
import gymnasium as gym
env = gym.make("stable_gym:HalfCheetahCost-v1")

state

The current system state.

Type:

numpy.ndarray

dt

The environment step size. Also available as tau.

Type:

float

reference_forward_velocity

The forward velocity that the agent should try to track.

Type:

float

Initialise a new HalfCheetahCost environment instance.

Parameters:

• reference_forward_velocity (float, optional) – The forward velocity that the agent should try to track. Defaults to 1.0.

• randomise_reference_forward_velocity (bool, optional) – Whether to randomize the reference forward velocity. Defaults to False.

• randomise_reference_forward_velocity_range (tuple, optional) – The range of the random reference forward velocity. Defaults to (0.5, 1.5).

• forward_velocity_weight (float, optional) – The weight used to scale the forward velocity error. Defaults to 1.0.

• include_ctrl_cost (bool, optional) – Whether you also want to penalize the half cheetah if it takes actions that are too large. Defaults to False.

• ctrl_cost_weight (float, optional) – The weight used to scale the control cost. Defaults to 1e-4.

• reset_noise_scale (float, optional) – Scale of random perturbations of the initial position and velocity. Defaults to 0.1.

• exclude_current_positions_from_observation (bool, optional) – Whether to omit the x- and y-coordinates of the front tip from observations. Excluding the position can serve as an inductive bias to induce position-agnostic behaviour in policies. Defaults to True.

• exclude_reference_from_observation (bool, optional) – Whether the reference should be excluded from the observation. Defaults to False.

• exclude_reference_error_from_observation (bool, optional) – Whether the error should be excluded from the observation. Defaults to True.

• exclude_x_velocity_from_observation (bool, optional) – Whether to omit the x- component of the velocity from observations. Defaults to False.

• action_space_dtype (union[numpy.dtype, str], optional) – The data type of the action space. Defaults to np.float32.

• observation_space_dtype (union[numpy.dtype, str], optional) – The data type of the observation space. Defaults to np.float64.

• **kwargs – Extra keyword arguments to pass to the HalfCheetahEnv class.

reference_forward_velocity

_randomise_reference_forward_velocity

_randomise_reference_forward_velocity_range

_forward_velocity_weight

_include_ctrl_cost

_exclude_reference_from_observation

_exclude_reference_error_from_observation

_exclude_x_velocity_from_observation

_action_space_dtype

_observation_space_dtype

_action_dtype_conversion_warning = False

state = None

low

high

observation_space

cost(x_velocity, ctrl_cost)

Compute the cost of a given x velocity and control cost.

Parameters:

• x_velocity (float) – The HalfCheetah’s x velocity.

• ctrl_cost (float) – The control cost.

Returns:

tuple containing:

• cost (float): The cost of the action.

• info (dict): Additional information about the cost.

Return type:

(tuple)

step(action)

Take step into the environment.

Note

This method overrides the step() method such that the new cost function is used.

Parameters:

action (np.ndarray) – Action to take in the environment.

Returns:

tuple containing:

• obs (np.ndarray): Environment observation.

• cost (float): Cost of the action.

• terminated (bool): Whether the episode is terminated.

• truncated (bool): Whether the episode was truncated. This value is set by wrappers when for example a time limit is reached or the agent goes out of bounds.

• info (dict): Additional information about the environment.

Return type:

(tuple)

reset(seed=None, options=None)

Reset gymnasium environment.

Parameters:

• seed (int, optional) – A random seed for the environment. By default None.

• options (dict, optional) – A dictionary containing additional options for resetting the environment. By default None. Not used in this environment.

Returns:

tuple containing:

• obs (numpy.ndarray): Initial environment observation.

• info (dict): Dictionary containing additional information.

Return type:

(tuple)

property tau

Alias for the environment step size. Done for compatibility with the

other gymnasium environments.

property t

Environment time.

property physics_time

Returns the physics time.