Abstract
We present an algorithm for skill discovery from expert demonstrations. The algorithm utilizes Large Language Models (LLMs) to first generate an initial segmentation of the trajectories. Following that, our proposed hierarchical variational inference framework incorporates additional information generated from the LLM to discover reusable skills by merging trajectory segments. Additionally, the algorithm introduces a novel auxiliary objective based on the Minimum Description Length principle that helps guide this skill discovery process. Our results demonstrate that agents learned this way discover skills that help accelerate learning on new long-horizon tasks in BabyAI, a gridworld navigation environment, as well as ALFRED, a household simulation environment.
