Robotics and Physical AI

Data-efficient, reliable vision-language-action (VLA) models for real-world humanoid robots — covering data collection, model development, and deployment. Physical AI is the wider thread: embodied systems that perceive, reason about, and act in the physical world with the reliability that real hardware demands.

Project Highlights

Vision-Language-Action Models for Humanoid Robots

Development of data-efficient, reliable VLA models for real-world humanoid robot use cases at Samsung SDS, spanning the full path from data collection through model development to deployment on physical platforms.

Embodied Data Collection & Curation

Pipelines for gathering and curating demonstration data for embodied learning, with an emphasis on getting strong policies out of modest data budgets rather than scaling collection indefinitely.

Grounded Multimodal Perception

Grounding language instructions and visual observations in 3D scene understanding, drawing on multi-view 3D semantic segmentation and self-supervised representation learning for geometric data.

Reliability & Uncertainty for Embodied Policies

Quantifying policy uncertainty so an embodied system can tell a confident action from a guess — carrying the Bayesian and uncertainty-aware methods developed for vision and language models into control.

Deployment on Real Hardware

Closing the gap between benchmark performance and behavior on physical robots, where sensing noise, latency, and long-tail conditions determine whether a policy is actually usable.

Modeling Physical Systems

Probabilistic and Gaussian Process modeling of physical processes and equipment, including uncertainty quantification and meta-modeling of asset digital twins — earlier work that informs how I approach physical-world models today.

Related Publications

The robotics work at Samsung SDS is largely unpublished. The publications below are the methodological foundations it builds on — uncertainty estimation, data-efficient 3D perception, and probabilistic modeling of physical systems.

Bayesian Active Learning for Semantic Segmentation
arXiv:2408.01694, 2024

Efficient uncertainty estimation for dense scene understanding — the perception side of embodied systems that must label and reason about their surroundings under a limited annotation budget.

ICML 2024 Improving Instruction Following in Language Models through Proxy-Based Uncertainty Estimation
International Conference on Machine Learning (ICML), 2024

Uncertainty-aware reward modeling for instruction following — directly relevant to VLA models, which must follow language instructions and know when their interpretation is uncertain.

Self-Supervised Contrastive Representation Learning for 3D Mesh Segmentation
arXiv:2208.04278, 2022

Learning useful representations of 3D geometry without manual annotation, a building block for spatial understanding in embodied settings.

Modeling and Optimizing the Impact of Process and Equipment Parameters in Sputtering Deposition Systems Using a Gaussian Process Machine Learning Framework
IEEE Transactions on Semiconductor Manufacturing, 2021

Gaussian Process modeling and optimization of a real physical deposition process, in collaboration with MIT — probabilistic modeling of physical systems with calibrated uncertainty.

Full publication list on Google Scholar →