Heejun Yoon

Welcome! 👋

Hello! I’m Heejun Yoon, an post-master researcher at the Korea Institute of Science and Technology (KIST).

My ultimate goal is to build application-driven AI to tackle our planet’s most pressing environmental and conservation challenges.

To achieve this, my work focuses on building intelligent systems that can truly perceive, understand, and act in our complex physical world. This research—bridging the gap between digital models and physical reality—is the essential foundation for creating AI that can operate reliably in the wild.

I hold M.S. and B.S. degrees in Electrical Engineering from Ewha Womans University, and I minored in Mathematics as an undergraduate.

💡 My Current Research

At KIST, I’m focused on developing the core capabilities for intelligent agents. My research is currently centered on two main areas:

• Multiview Scene Understanding: How can an AI understand a 3D space from multiple, limited 2D viewpoints? I’m developing Vision-Language Models (VLMs) that integrate these views to build a rich, contextual understanding of a scene.
• Embodied AI (Navigation & Planning): Taking understanding a step further—how can an agent act on that understanding? I’m exploring how Large Language Models (LLMs) and Reinforcement Learning (RL) can work together for complex task planning and navigation in simulated worlds like Isaac Sim and Omnigibson.

🌱 My Research Vision: The Questions That Drive Me

My current research at KIST builds the ‘how’—the technical skills to make an agent perceive, understand, and act. But my ‘why’ comes from a deep desire to aim this technology at our planet’s most pressing challenges.

My vision is to develop “Application-Driven AI”—moving beyond generic benchmarks to build models that solve specific, real-world constraints. This vision is built on tackling one core problem and pursuing two fundamental research pillars:

The Core Challenge: AI Beyond Web-Scale Data

Most modern AI excels on standardized benchmarks, but these models are ill-suited for high-stakes environmental problems. Ecological data is the opposite: it’s sparse, expensive to collect, geographically biased, and incredibly messy. This is the central problem I am focused on solving.

Pillar 1: Building a Unified Picture from Fragmented Data

To understand an ecosystem, we must connect satellite images, on-the-ground acoustic sensors, drone footage, and even textual policy reports.

• Question : How can we build robust multimodal models that fuse these diverse and imperfect data streams into a single, reliable understanding to inform action?

Pillar 2: Optimizing Decisions When the “Cost of Error” is Real

In conservation, a model’s mistake isn’t just a lower accuracy score; it’s a habitat we fail to protect or resources wasted.

• Question : How can we build decision-making frameworks that move beyond simple “error rates” and help policymakers weigh the true real-world cost of their decisions?

My ultimate goal is to create tools that are not just technically novel, but are trustworthy, realistic, and truly valuable for the stakeholders, scientists, and policymakers on the front lines.

🚀 My Journey So Far

My path has been a progressive journey of understanding signals and systems.

I started with 1D radar signals in undergrad, learning to detect motion and vital signs. My master’s research expanded to 2D machine vision, where I built models to find tiny defects in semiconductors.

Now, my work at KIST has moved into 3D scenes and embodied action. This progression—from 1D signals to 2D images, and now to 3D interaction—has given me a powerful toolkit to tackle complex, real-world data.

My Mission 🌱

Through my work, I aim to make complex data accessible and actionable and push the boundaries of AI practices. I firmly believe that technology should be developed responsibly and used for the greater good.

If you’re curious about my research, upcoming projects, or just want to connect, feel free to contact via e-mail or LinkdIn(heejun-yoon). Looking forward to sharing my work and insights with you!