OpenAI Gym Car Race

An OpenAI Gym environment for simulating a self driving car on a race track. Built for NYU SelfDrive.

This project is still under active development. Bug reports are welcome.

Prerequisites

Ensure that you are using python3 and that all required python modules are installed by running the following:

pip install --no-cache-dir -r requirements.txt

Running

Run python example.py in the root of this repository to execute the example project. Doing so will create the necessary folders and begin the process of training a simple nueral network. After training has completed, a window will open showing the car navigating the pre-saved track using the trained model.

To create a new track, pass in new=True to the reset function of the environment. Doing so will open a window inside of which you have the following options:

• Left click: remove a block
• Right click: put back a block
• S key: put down a green start block
• F key: put down a red finish block
• ESC key: finish creating the track and run the simulationn

The track created will be saved in whatever file was specified under track_file in config.py or any other overriding configuration.

If you choose to run the simulation without rendering the car, you can omit env.render() from the main loop in your project.

Configuring

Much of the general configuration can be done through editing the values in config.py. A custom reward function can also be passed into the environment by defining it like the following example and adding a reference to it in the reward.function field in the config file:

@Car.reward_function
def my_reward_func(car):
    # Function must take an instance of the car as a paramter
    ...
    # Calculate reward here using values from car

    return reward

# Make sure to add my_reward_func to either config.py or an
# overidden config

car = Car()
env.add_car(car)

If no reward function is provided, a default implementation is used using the corresponding values in config.py.

For an example on how to specify and train multiple configurations at the same time, see test_bench.py.

Logging

Logging is handled automatically by Tensorboard. To view the logged data and visualizations, run tensorboard dev upload --logdir './my_log_dir' where my_log_dir is directory of the logs you want to view. Navigate to url from the output of the above command to see the graphs.