SEONVIN CHO

Research

Research Interests

Offline Reinforcement Learning Sequential Decision Making Generative Models

Recent Work

Role-Adaptive Policy Optimization for Offline Reinforcement Learning

Seonvin Cho, Soohyun Choi, Songnam Hong†
Preprint · †Corresponding author

  • Proposed Role-Adaptive Policy Optimization (RAPO), which learns policy-update coefficients according to whether actions are used for critic bootstrapping or for execution
  • Separated bootstrap and execution actors in TD3+BC and adapted only the inverse temperature for IQL, improving both base algorithms on D4RL locomotion and AntMaze
arXiv

Is One Step Enough for Offline Policy Improvement?

Soohyun Choi*, Seonvin Cho*, Songnam Hong†
Preprint · *Equal contribution · †Corresponding author

  • Studied multi-step proximal policy improvement that re-centers each stage on the preceding policy, asking when a single improvement step is enough
  • Distinguished subdivision of a fixed total horizon from additional refinement, with experiments on TD3+BC and IQL-based policy extraction in D4RL locomotion
code arXiv

PathBridger: Subgoal Bridges for Offline Goal-Conditioned Reinforcement Learning

Soohyun Choi*, Seonvin Cho*, Songnam Hong†
Preprint · *Equal contribution · †Corresponding author

  • Developed a hierarchical offline GCRL framework that connects subgoal selection to short-horizon execution through state-space bridges
  • Combined generative endpoint proposals, transitive value-based selection, and inverse-dynamics decoding to produce executable action chunks from offline trajectories
code arXiv
Overview of the PathBridger bridge policy

Ongoing Research

Extensions of PathBridger

  • Skill Learning and Composition: Investigating how to learn reusable skills from offline trajectories and compose them for long-horizon goal-conditioned tasks.
  • Latent-Space Bridge Planning: Exploring learned representations and geometrically structured latent spaces for goal-conditioned bridge generation and short-horizon execution.

Research Topics

Policy Optimization in Offline Reinforcement Learning

  • Treat behavior-regularized policy updates as proximal steps along a critic-induced gradient flow
  • Use this view to reason about multi-step refinement under different policy geometries and imperfect critic estimates

Generative Models for Sequential Decision Making

  • Study generative approaches to sequential decision making, including sequence models, diffusion models, energy-based methods, and flow-based models
  • Use them for trajectory generation, planning, and long-horizon control