Wonjun Jo

I am a Ph.D. candidate in the Department of Electrical Engineering at POSTECH. I conduct my research at the Advanced Machine Intelligence (AMI) Lab at KAIST, advised by Prof. Tae-Hyun Oh.

I am interested in Self-supervised Learning and Multi-modal Learning for preserving and improving the generalization capability of World Models, toward robust Embodied AI and real-world robotics.

More specifically, my recent research focuses on Vision-Language-Action (VLA) and World Action Models (WAM) for robotics, with additional interest in sound and tactile signals for holistic world understanding.

Embodied AI Self-supervised Learning Multi-modal Learning World Models

Email | CV | Scholar | LinkedIn

Research

	When Fine-Tuning Drifts: Mitigating Policy Drift with Base Action-Field Regularization Wonjun Jo, Nam Hyeon-Woo, Yohan Park, Hyunwoo Ha, Tae-Hyun Oh Under-review Project Page \| Paper Mitigating policy drift for robust few-shot fine-tuning of pretrained robot policies.
	LiDAR-Anchored Collaborative Distillation for Robust 2D Representations Wonjun Jo, Hyunwoo Ha, Kim Ji-Yeon, Hawook Jeong, Tae-Hyun Oh Under-review Project Page \| Paper Improving the robustness and generalization of self-supervised visual representations with LiDAR guidance.
	DarkQA: Benchmarking Vision-Language Models on Visual-Primitive Question Answering in Low-Light Indoor Scenes Yohan Park, Hyunwoo Ha, Wonjun Jo, Tae-Hyun Oh IEEE Robotics and Automation Letters (RA-L), 2026 Project Page \| Paper Benchmarking the robustness of embodied vision-language models under low-light conditions.
	Self-Supervised Collaborative Distillation: Enhancing Lighting Robustness and 3D Awareness Wonjun Jo, Hyunwoo Ha, Kim Ji-Yeon, Hawook Jeong, Tae-Hyun Oh Workshop on Wild3D, IEEE/CVF International Conference on Computer Vision (ICCV), 2025 Paper Improving pretrained visual representations for lighting robustness and 3D-aware generalization.
	The Devil is in the Details: Simple Remedies for Image-to-LiDAR Representation Learning Wonjun Jo, Kwon Byung-Ki, Kim Ji-Yeon, Hawook Jeong, Kyungdon Joo, Tae-Hyun Oh Asian Conference on Computer Vision (ACCV), 2024 Project Page \| Paper Improving self-supervised image-to-LiDAR representation learning for robust 3D perception.

Notes

Research Core

Self-Supervised Learning

Multi-Modal Learning

World Models

Embodied AI

Energy-Based Models (Yann LeCun)

Free Energy Principle (Karl Friston)

Music

No Time for Caution Hans Zimmer YouTube
Can You Hear The Music Ludwig Goransson YouTube
Time Hans Zimmer YouTube
We Have to Go Steve Jablonsky YouTube
Cloud Atlas End Title Tom Tykwer, MDR Rundfunkchor, MDR Sinfonieorchester YouTube
Tears in the Rain Hans Zimmer, Benjamin Wallfisch YouTube
Seven Worlds One Planet Suite Hans Zimmer YouTube
Contact - End Credits Alan Silvestri YouTube
Transformation (End Titles) Hans Zimmer YouTube
Leaving Caladan Hans Zimmer YouTube
You Don't Have To Kris Bowers YouTube
F1 Hans Zimmer YouTube
Believe in the Hail Mary Daniel Pemberton YouTube