Junwon Seo

I am a PhD student in the Robotics Institute at Carnegie Mellon University (CMU), advised by Prof. Andrea Bajcsy.

My research focuses on developing uncertainty-aware robotic systems that ensure safety and robustness in complex, open-world environments. I study the theoretical foundations and algorithms for quantifying uncertainty in learning-based robotic systems and leverage control-theoretic and information-theoretic principles to mitigate it.

Before joining CMU, I was a research officer at the Defense AI Center of the Agency for Defense Development (ADD), where I researched uncertainty-aware navigation in complex and unknown environments. I completed my bachelor's degree (Summa Cum Laude) in Computer Science and Engineering at the Seoul National University (SNU).

Email  |  CV  |  GitHub  |  Google Scholar  |  YouTube  |  LinkedIn

• 

  [Nov 2025]

  One paper accepted to WACV 2026! OW-Rep is an open world object detection framework that jointly detects unknown objects and learns instance embeddings, enabling the detector to capture fine-grained semantic relationships between instances that are useful for downstream tasks.
• 

  [Sep 2025]

  New paper on arXiv about constraint-parameterized latent safety filters that can adapt to user-specified safety constraints at runtime (AnySafe).
• 

  [Jul 2025]

  One paper accepted to CoRL 2025! UNISafe is an uncertainty-aware latent safety filter that unifies reachability analysis in a latent world model with OOD detection to proactively steer robots away from both known and out-of-distribution failures.
• [Sep 2024] I started my PhD in the Robotics Institute at Carnegie Mellon University (CMU), happily advised by Prof. Andrea Bajcsy.
• [Jul 2024] Two papers accepted to IROS 2024! One is using evidential deep learning for uncertainty-aware semantic mapping (selected as Best Cognitive Paper Finalist), and the other is using meta-learning for online adaptation of traversability cost maps for off-road navigation.
• [May 2024] I finished my military service in Agency for Defense Development (ADD) as a Research Officer for National Defense (ROND), where I researched uncertainty-aware navigation in unstructured and unknown environments. Junwon is free!
• [May 2024] Presented three papers at ICRA 2024! Our RA-L paper introduces self-supervised traversability estimation for off-road environments; another work explores domain adaptive object detection in real-world adverse weather; and a workshop paper proposes uncertainty-aware semantic mapping using Dempster-Shafer theory of evidence.
• [Oct 2023] Presented my RA-L paper on learning vehicle-specific traversability with self-supervised learning at IROS 2023.
• [Mar 2023] Our paper on bridging active exploration and uncertainty-aware deployment accepted to RSS 2023!

* indicates equal contribution

Constraint-parameterized latent safety filters that can adapt to user-specified safety constraints at runtime.

Evidential Ellipsoidal Bayesian Kernel Inference for uncertainty-aware continuous 3D semantic mapping.

Uncertainty-aware latent safety filters that unify reachability analysis in a latent world model with OOD detection to prevent both known and unseen safety hazards.

Open World Object Detection framework that jointly detects unknown objects and learns instance embeddings.

Uncertainty-aware semantic BKI mapping framework for robust deployments in off-road environments using Evidential Deep Learning.

Meta-learning framework for learning a global model that accurately predicts off-road terrain traversability.

UFO: Uncertainty-aware LiDAR-image Fusion for Off-road Semantic Terrain Map Estimation.

Unsupervised domain adaptation for robust object detection in real-world adverse weather conditions.

Active exploration for learning vehicle dynamics model and uncertainty-aware deployment.

Visual terrain traversability learning in off-road that utilizes only self-supervision without human annotations.

LiDAR-based self-supervised traversability estimation framework that can be easily scalable to various vehicles and sensor types.

Handling uncertainty problems in self-supervised traversability estimation with metric learning.

Perception and control algorithms for off-road navigation: learning-based traversability cost map, semantic terrain classification map, moving object detection and tracking, and vehicle dynamics learning.

Built an autonomous driving dataset with multi-spectral images and LiDAR in adverse weather conditions and developed a robust traversable area detection and object detection algorithm.

Developed a learning-based traversability map for autonomous robotic exploration under subterranean environments.

The light stage system and face normal map reconstruction algorithm for 3D face reconstruction with high fidelity.

A Structured-light 3D scanner for high-quality hair and face geometry reconstruction.

Website template adapted from Jon Barron's website