Autonomous driving via reinforcement learning in the CARLA simulator

In this project reinforcement is used to teach a car how drive a path defined by a series of checkpoints. The CARLA simulator is used as the environment for the car.


The list of checkpoints on the map 'Town02'. Blue is start/ end.

The final result of the trained model can be seen in the following Youtube Video.

Prerequisites

The program was written with Python 3.7 using the 9.11 version of the CARLA Simulator. All needed modules can be installed running the run.py script.

Running the program

The program expects to be nested in the PythonAPI folder in the CARLA folder.

Training a model

Run run.py to install all needed modules and start the training process. The attribute load_model_name in main.py can be edited to load an existing model. The value of None will create new model.

Testing a model

Run the test_model.py script to test a model on the defined track. By editing the load_model_name in this script, a specific model can be tested. Testing the model for overfitting by going around the track backwards can be done by changing the REVERSE flag from False to True.

Sources