Hanbyel Cho

I am currently an AI Robotics Research Scientist at Samsung Electronics, working on AI-driven humanoid whole-body control and locomotion. My research integrates human motion understanding, reinforcement learning, and control theory to enable humanoid robots to learn and adapt from human motion in real-world environments.

I earned my PhD in Electrical Engineering from KAIST, where I was advised by Prof. Junmo Kim. My doctoral research focused on capturing and reconstructing human-related subjects—such as body pose and shape—under real-world conditions for practical applications in immersive environments.

During my PhD studies, I gained valuable experience through two internships at Meta Reality Labs—one in Pittsburgh, PA, USA (2023), and another in Redmond, WA, USA (2024). I was also honored to be selected as a finalist for the Qualcomm Innovation Fellowship in both 2022 and 2023.

If you're interested in collaborating, feel free to reach out!

tlrl4658 [at] gmail.com /  CV (updated: Feb 2025)  /  Bio  /  Scholar  /  Linkedin  

Publications

My current research centers on learning-based humanoid whole-body control and locomotion, aiming to enable robots to learn agile and adaptive motion from data. Previously, I studied 3D human reconstruction and motion understanding, which naturally led to my current interest in AI-driven humanoid behavior learning. Below are my recent publications; some papers are highlighted.

A lightweight and hyperparameter-free debiasing method that whitens feature representations to remove bias, achieving fairness and utility trade-off without adversarial learning.

Efficient 4D dynamic scene editing method using 4D Gaussian Splatting, focusing on static 3D Gaussians and score distillation refinement to achieve faster, high-quality edits.

Proposes a novel loss-aware vulnerability proxy (LAVP) for gradient inversion attacks in federated learning, using the maximum or minimum eigenvalue of the Hessian to capture sample vulnerabilities beyond traditional gradient norm.

Proposes a novel generative framework for 3D human mesh recovery, leveraging denoising diffusion to model multiple plausible outcomes and address the inherent ambiguity in the task.

Leveraging neural feature fields to render multi-view feature maps and enforcing cross-view consistency enables accurate 3D human mesh recovery from a single image.

Replacing naive GRU-based modeling with a Vision Transformer and a learnable Channel Rearranging Matrix reduces motion fragmentation, boosting human mesh recovery accuracy, robustness, and efficiency.

Masked face recognition approach that leverages multi-focal spatial attention to precisely isolate unmasked features.

Replacing softmax in non-local blocks with a simple scaling factor mitigates its over-reliance on vector magnitude, thereby improving performance, robustness, and efficiency.

3D human pose estimation model that leverages MAML and synthetic distorted data to rapidly adapt to various camera distortions without requiring calibration.

Introduces a novel Convolutional Unit that aligns whitening theory with convolutional architectures, significantly boosting the stability and performance of Batch Whitening.

Stochastic modeling-based face super-resolution method that separates deterministic and stochastic attributes to reduce uncertainty and improve reconstruction quality.