carracing_obstacles

Variant of OpenAI Gym CarRacing environment with static obstacles

Environment Description

This environment is a modified version of the CarRacing environment from OpenAI Gym, and is also inspired by NotAnyMike's CarRacing variant as well here. It has the following features:

• Obstacles are randomly placed in a particular road tile with probability OBSTACLE_PROB, and are randomly placed on either the left or right side of the road.
• Obstacles are spaced apart by a minimum of OBSTACLE_SPACING tiles.
• Obstacles incur a fixed penalty (given by OBSTACLE_PENALTY)

Requirements

• Gym 0.21.0

Usage

To test the environment using human keyboard inputs, simply run:

python car_racing_obstacles.py

To use the environment in OpenAI Gym RL scenarios, move car_racing_obstacles.py to your working directory (or add the directory containing this repo to your PYTHONPATH) and do the following:

from car_racing_obstacles import CarRacingObstacles
from gym.wrappers.time_limit import TimeLimit
env = TimeLimit(CarRacingObstacles(),max_episode_steps=1000)
...

By default, OBSTACLE_PROB will be set to 0.05. Note that the time limit of 1000 timesteps is to ensure that behavior is identical to that of the built-in CarRacing-v0 environment.

When we call env.step(action), the returned info dictionary includes the following keys:

• num_obstacles: total number of obstacles in the track
• num_collisions: total number of collisions with obstacles (note that we can collide with the same obstacle more than once)