Unsupervised Hierarchical Skill Discovery

TL;DR

HiSD presents an unsupervised framework for discovering hierarchical skill structures directly from observational data by coupling temporal action segmentation with grammar-based sequence compression. It uses unbalanced optimal transport for robust skill segmentation and Sequitur grammar induction to build a global, reusable hierarchy over discovered skills, all without action or reward supervision. Evaluations in Craftax and Minecraft show HiSD delivers more accurate skill segmentation and deeper, more coherent hierarchies than baselines, and that these abstractions can accelerate downstream reinforcement learning through modular options and initiation/termination models. The work demonstrates a scalable path to learning structured, compositional representations from unlabelled data, with practical impact for enabling efficient learning in high-dimensional domains.

Abstract

We consider the problem of unsupervised skill segmentation and hierarchical structure discovery in reinforcement learning. While recent approaches have sought to segment trajectories into reusable skills or options, most rely on action labels, rewards, or handcrafted annotations, limiting their applicability. We propose a method that segments unlabelled trajectories into skills and induces a hierarchical structure over them using a grammar-based approach. The resulting hierarchy captures both low-level behaviours and their composition into higher-level skills. We evaluate our approach in high-dimensional, pixel-based environments, including Craftax and the full, unmodified version of Minecraft. Using metrics for skill segmentation, reuse, and hierarchy quality, we find that our method consistently produces more structured and semantically meaningful hierarchies than existing baselines. Furthermore, as a proof of concept for utility, we demonstrate that these discovered hierarchies accelerate and stabilise learning on downstream reinforcement learning tasks.

Figures (17)

• Figure 1: Example of HiSD applied to a Minecraft trajectory. Step 1: HiSD segments the observational trajectory into distinct skills, such as get wood, craft tools, and gather stone. Step 2: Using these segmented trajectories, HiSD applies a grammar-based compression algorithm to induce a hierarchy over the discovered skills, revealing reusable subroutines and their temporal organisation.
• Figure 2: Overview of the HiSD pipeline. Demonstration trajectories are first segmented into skills. These skill sequences are then compressed and structured using a modified Sequitur algorithm, which identifies recurring mid-level subroutines across the dataset. The resulting grammar defines a hierarchical task decomposition.
• Figure 3: Example of the skill segmentation performance in the Stone Pickaxe Static Task in Craftax for all three baselines. Colours indicate discovered skills: wood, table, wooden_pickaxe, stone, and stone_pickaxe.
• Figure 4: Example of the skill segmentation performance in the Minecraft Mapped Task for all three approaches. Colours indicate discovered skills: Walk, Mine Log, Craft Planks, Craft Table, Craft Stick, Use Table, Craft Wooden Pickaxe, Mine Table, Mine Grass, Mine Dirt, and Mine Stone.
• Figure 5: Mean episode rewards (±1 SD) over environment time steps for 10 random seeds on the Craftax Wooden Pickaxe task. The HiSD hierarchy (orange) achieves higher and more stable performance than both the OMPN hierarchy (green) and the Skills-Only (HiSD) variant (purple), while closely matching the ground truth Hierarchy (blue). Primitive action PPO (cyan) fails to solve the task.