This repository contains the implemetations for the following:
- Internal state-based car control.
- Self driving in CARLA's Town02 map.
The reinforcement learning method used for self driving was the Deep Q-Network. The Xception model was used as the DQN neural network.
The car has a camera attached to the hood. The frames captured by the camera are fed into the neural network for training in batches. A collision sensor is also attached to the vehicle. It is used in tandem with the episode duration to determine the reward.
The highest observed reward for the was around -130 at episode 100. The lowest observed reward was -340 at episode 1.