Devising Overtraining Curriculum Learning Techniques to Enhance Autonomous Navigation Reinforcement Learning

Autonomous navigation, which involves independently navigating an agent through obstacles, is commonly approached through reinforcement learning (RL). However, RL is notoriously difficult to train, which hinders model performance. So, this paper proposes “overtraining” curriculum learning (CL) techniques to enhance training processes. Inspired by human phenomena, overtraining trains an agent in harder environments than those it is tested in, allowing for expedited learning. To validate this, a car agent was trained using RL to navigate a series of obstacles in Python. Various “overtraining” approaches, alongside existing end-to-end and CL techniques, were benchmarked. After measuring performance across CL magnitudes and environment difficulties, the static overtraining approach demonstrated 9.05% more reward than the baseline. This provides an easy-to-use method of enhancing autonomous navigation algorithms, enabling innovations in self-driving, indoor navigation, robotics, and more.