Reinforcement learning in a driving simulator (CARLA)

The client's domain is confidential, so this stays general. The task was to train an agent to complete a set of objectives inside CARLA, a realistic driving simulator, using what its camera sees to decide what to do.

I wrapped the simulator as an OpenAI Gym environment, processed the camera images with OpenCV, and trained the agent with TensorFlow and TF-Agents. Most of the work was in the reward: I refined it over many runs until the agent learned the behavior we wanted rather than a shortcut, alongside systematic hyperparameter tuning and exploration strategies such as epsilon-greedy and UCB.