Parking

A goal-conditioned continuous control task in which the ego-vehicle must park in a given space with the appropriate heading.

Usage

env = gym.make("parking-v0")

Versions

ID	Description
parking-v0	Default parking task with an empty lot.
parking-ActionRepeat-v0	Lower policy frequency (1 Hz) and shorter episode (20 s).
parking-parked-v0	Lot populated with 10 parked vehicles as obstacles.

Default configuration

{
    "observation": {
        "type": "KinematicsGoal",
        "features": ['x', 'y', 'vx', 'vy', 'cos_h', 'sin_h'],
        "scales": [100, 100, 5, 5, 1, 1],
        "normalize": False
    },
    "action": {
        "type": "ContinuousAction"
    },
    "reward_weights": [1, 0.3, 0, 0, 0.02, 0.02],
    "success_goal_reward": 0.12,
    "collision_reward": -5,
    "steering_range": np.deg2rad(45),
    "simulation_frequency": 15,
    "policy_frequency": 5,
    "duration": 100,
    "controlled_vehicles": 1,
    "vehicles_count": 0,
    "add_walls": True,
    "screen_width": 600,
    "screen_height": 300,
    "centering_position": [0.5, 0.5],
    "scaling": 7,
    "show_trajectories": False,
    "render_agent": True,
    "offscreen_rendering": None
}

More specifically, it is defined in:

classmethod ParkingEnv.default_config() → dict

Default environment configuration.

Can be overloaded in environment implementations, or by calling configure(). :return: a configuration dict

API

class highway_env.envs.parking_env.ParkingEnv(config: dict = None, render_mode: str | None = None)

A continuous control environment.

It implements a reach-type task, where the agent observes their position and speed and must control their acceleration and steering so as to reach a given goal.

Credits to Munir Jojo-Verge for the idea and initial implementation.

classmethod default_config() → dict

Default environment configuration.

Can be overloaded in environment implementations, or by calling configure(). :return: a configuration dict

define_spaces() → None

Set the types and spaces of observation and action from config.

compute_reward(achieved_goal: ndarray, desired_goal: ndarray, info: dict, p: float = 0.5) → float

Proximity to the goal is rewarded

We use a weighted p-norm

Parameters:

• achieved_goal – the goal that was achieved

• desired_goal – the goal that was desired

• info (dict) – any supplementary information

• p – the Lp^p norm used in the reward. Use p<1 to have high kurtosis for rewards in [0, 1]

Returns:

the corresponding reward