Skill Learning Using a Single Demonstration for Hierarchical Robotic Navigation

Training robots to perform complex navigation tasks traditionally requires extensive datasets of expert demonstrations, which are expensive and time-consuming to collect. This paper introduces a hierarchical reinforcement learning approach that distils meaningful skills using a single expert trajectory. By combining reverse curriculum learning and goal-conditioned reinforcement learning, our method segments complex tasks into manageable sub-tasks and learns these sub-tasks using sparse rewards. A learned skill selection mechanism then dynamically chains the skills based on their estimated values. We evaluate our approach in a high-dimensional autonomous driving scenario, demonstrating skill acquisition and task completion rates that surpass a state-of-the-art baseline that also uses a single demonstration. Our results show that flexible skill extraction and chaining using a reduced amount of data can effectively tackle the challenges of sparse rewards and planning in a robotic navigation environment.