Seohong Park

Papers

Total Citations

H-Index

About

Seohong Park is a leading researcher in unsupervised skill discovery and hierarchical reinforcement learning, whose work is fundamentally reshaping how autonomous agents learn diverse, reusable behaviors without external rewards. Park’s core contributions center on developing theoretically grounded frameworks that move beyond naive mutual information maximization. In their highly influential 2022 work, "Lipschitz-constrained Unsupervised Skill Discovery" (8 citations), Park identified a critical failure mode in prior methods—their tendency to produce trivial, easily distinguishable skills—and introduced a Lipschitz constraint that forces agents to learn genuinely distinct and behaviorally meaningful skills. Building on this, their 2023 paper "Controllability-Aware Unsupervised Skill Discovery" (8 citations) pioneered a paradigm shift by incentivizing agents to seek out skills that grant them greater control over their environment, rather than settling for simple, low-effort behaviors. Most recently, in "GHIL-Glue: Hierarchical Control with Filtered Subgoal Images" (2025), Park is extending these principles into the visual domain, leveraging pretrained generative models to produce robust subgoals for hierarchical robot learning. With a growing citation footprint and a clear trajectory from foundational theory to practical robotics, Park is establishing a new standard for how machines can autonomously structure their own learning curricula.