Chung-Yeon Lee

I'm a research scientist at Surromind in Seoul, where I lead a small research team focused on multimodal, physical, and generative AI models and their applications across various domains. At Surromind, I've worked on sound- and vibration-based fault diagnosis systems, service and field robot learning, and generative models (LLM, VLM, VLA), among others. I did my PhD at Seoul National University, where I was advised by Byoung-Tak Zhang.

Email  /  CV  /  Scholar  /  Github

Research

I'm interested in bridging AI and robots, aiming to build intelligent physical agents capable of spatial reasoning, high-level planning, and adaptive interaction in complex real-world environments.

A multimodal attention-based neural representation model that integrates RGB, depth, and semantic information to enable generalizable spatial representations and novel-view rendering across diverse indoor environments.

A deep autoencoder framework combining RGB, depth, audio, and force-torque sensor data to detect object slip anomalies in mobile manipulation robots by learning latent representations of normal conditions and identifying deviations in real-world noisy environments.

ROS-based visual perception framework that integrates various deep learning methods for object recognition, person identification, human pose estimation, scene captioning, and object-aware navigation.

A spatial perception framework for autonomous robots that enhances traditional geometric scene representations by integrating semantic object-aware features, thereby improving map generation, reducing tracking failures, and enabling more reliable place recognition in both home environments and large-scale indoor datasets like ScanNet.

Team Tidyboy is a robotics challenge team dedicated to solving home service tasks by integrating multiple AI models, including verbal and nonverbal human-robot interaction, 3D perception of known and unknown objects, and manipulation of various items through coordinated use of both the manipulator and the robot’s base.

Team Aupair aims to develop intelligent mobile cognitive robots with a machine learning. We envision a new paradigm of service robot with state-of-the-art deep learning methods to carry out difficult and complex real world tasks. We propose a novel integrated perception system for service robots which provides an elastic parallel pipeline to integrate multimodal vision modules, including state-of-the- art deep learning models. On top of that, we deploy highly sophisticated modules such as socially-aware navigation, visual question-answering, and schedule learning.

We propose a dual-memory deep learning architecture for lifelong learning of everyday human behaviors from non-stationary data streams, combining a slow-changing global memory with a fast-adapting local memory, and demonstrate its effectiveness on real-world datasets including image streams and lifelogs collected over extended periods.

Here we investigate the information flow network of the human brain during episodic memory retrieval. We have estimated local oscillation amplitudes and asymmetric inter-areal synchronization from EEG recordings in individual cortical anatomy by using source reconstruction techniques and effective connectivity methods during episodic memory retrieval. The strength and spectro-anatomical patterns of these inter-areal interactions in sub-second time-scales reveal that the episodic memory retrieval involves the increase of information flow and densely interconnected networks between the prefrontal cortex, the medial temporal lobe, and some subregions of the parietal cortex. In this network, interestingly, the SFG acted as a hub, globally interconnected across broad brain regions

We demonstrate a memory experiment that employs audio-visual movies as naturalistic stimuli. EEG recorded during memory formation show that oscillatory activities in the theta frequency bands on the left parietal lobe, and gamma frequency bands on the temporal lobes are related to overall memory formation. Theta and gamma power of the frontal lobes, and gamma power of the occipital lobes were both increased during retrieval tasks.

Recognizing the type of place a person is in (e.g., classroom, hallway, restaurant, outdoor) using photographs taken by a smartphone camera combined with GPS information, by training a modified SVM classifier where image features (SIFT) are the inputs and GPS serves as weighting information.

Here, we propose multi-layer structural wound synthesis on a 3D face. The approach first defines the facial tissue depth map to measure details at various locations on the face. Each layer of skin in a wound image has been determined by hue-based segmentation. In addition, we have employed disparity parameters to realise 3D depth in order to make a wound model volumetric. Finally, we validate our methods using 3D wound simulation experiments.

We propose an automatic method for exaggeration of facial expressions from motion-captured data with a certain personality type. The exaggerated facial expressions are generated by using the exaggeration mapping (EM) that transforms facial motions into exaggerated motions. As all individuals do not have identical personalities, a conceptual mapping of the individual’s personality type for exaggerating facial expressions needs to be considered. The Myers–Briggs type indicator, which is a popular method for classifying personality types, is employed to define the personality-type-based EM. Further, we have experimentally validated the EM and simulations of facial expressions.

We propose a method to generate exaggerated facial expressions based on individual personal styles by applying Non-Negative Matrix Factorization (NMF) to decompose facial motion data, then enhancing expression intensity using a style-dependent exaggeration rate (SER). Experiments show that this approach effectively produces subtle, personalized exaggerations across various basic emotions such as happiness, sadness, and surprise.

We present a method of shoal motion with an effective leadership of autonomous virtual fish. Shoal motion is led by a leader and computed by five steering behavior vectors including cohesion, separation, velocity, escape and goal vectors. Through experiments, we demonstrate that a leader of the motion simulation has great effect in leadership and great accuracy to guide the group.

*Last update in September 2025