Curriculum Learning For Autonomous Vehicles

This study investigates how the sequence of training environments affects performance in simple driving tasks for an autonomous driving agent. By training agents solely through interaction with maps of varying difficulty, we demonstrate that transfer learning enhances performance within single-environment driving scenarios. However, we find that agents struggle to master advanced driving capabilities and fail to generalize well to new environments, regardless of the sequence of training data. We conclude by looking at areas to build on this work such by combining imitation learning with curriculum learning and developing curriculum-specific MDP.