Physical AI / World Models

Youngjoon Jeong

I am a Ph.D. candidate in the Graduate School of Data Science at Seoul National University, advised by Taesup Kim. My goal is to build embodied agents that generalize to unseen situations with minimal data and computation.

News

Sep 2026 PoLAR was accepted to CoRL 2026.
Jun 2026 Our team received support through the NVIDIA Academic Grant Program.
Jun 2026 Two preprints, PoLAR and J-PARC, are now available on arXiv.
Feb 2026 VILA was accepted to CVPR 2026.
Jan 2026 Sparse Imagination was accepted to ICLR 2026.

Research Interests

I study representation learning and world models to enable zero-shot generalization and efficient learning, planning, and control.

Generalization: enabling embodied agents to act in unseen situations, with an emphasis on zero-shot transfer across viewpoints, tasks, environments, and robot dynamics.
Efficiency: reducing the data and computation needed for robot policy learning and world-model planning through learned representations, including latent actions.

Research Experience

Ph.D. Researcher @ Seoul National University

Advisor: Taesup Kim

Seoul, South Korea Sep. 2023–Present

Selected projects

Latent actions for robot policy learning: VILA (CVPR 2026) and PoLAR (CoRL 2026) learn view-invariant and structured action representations from visual observations for downstream visuomotor policies.
Efficient visual world-model planning: Sparse Imagination (ICLR 2026) reduces the computational cost of planning with imagined future visual states.
Robust vision-language-action policies: J-PARC (arXiv 2026) studies joint-level physical faults and adaptive action correction under altered robot dynamics.