Research #

3D Scene Understanding for Mobile Robots #

Understanding the geometry and semantics of their surroundings is essential for autonomous robots to navigate and operate effectively in the physical world. How can robots derive such an understanding from 3D measurements such as LiDAR scans? My research addresses this question through scene classification, generative modeling, restoration, upsampling, and sim-to-real domain adaptation.

Sim2Real LiDAR segmentation with raydrop-aware diffusion sampling.

DRUM (ICRA 2026)

Fast LiDAR generation with rectified flows.

R2Flow (ICRA 2025)

LiDAR generation and completion with unconditional diffusion models.

R2DM (ICRA 2024)

Conditional diffusion for fast sparse-to-dense LiDAR upsampling.

LiDAR Upsampling (RO-MAN 2024)

Resolution-free GAN for LiDAR generation, restoration, and upsampling.

DUSty v2 (WACV 2023)

Raydrop-aware GAN for LiDAR generation.

DUSty (IROS 2021)

A multimodal LiDAR dataset for outdoor place categorization and scene understanding.

LiDAR Datasets (IJRR 2019)

Panoramic LiDAR place categorization using geometric and photometric CNN features.

Outdoor Place Categorization (Advanced Robotics 2018)

Vision-based Human Activity Recognition #

This work focuses on identifying people and understanding their activities in environments. The work includes gait recognition from RGB videos and LiDAR point clouds, with methods designed to handle changes in appearance, viewpoint, walking direction, measurement distance, and point-cloud density. It also explores multi-perspective visual lifelogging that combines first-, second-, and third-person views to describe human activities and interactions.

RGB gait recognition with style-feature sampling.

S2Gait (SII 2025)

Diffusion-based upsampling of sparse LiDAR gait sequences.

Gait Sequence Upsampling (SII 2025)

Disentangled gait feature swapping for robust RGB gait recognition.

Disentangled Gait Swapping (IEEE Access 2024)

Viewpoint- and distance-robust LiDAR gait recognition.

LiDAR-Based Gait Recognition (IEEE Access 2023)

LiDAR gait recognition robust to walking direction and measurement distance.

2V-Gait (SII 2022)

Identity-aware adversarial augmentation for gait representation learning.

Identity-Aware Augmentation (SII 2022)

Multi-view captioning from first-, second-, and third-person images.

Fourth-person visual lifelogging (ROBOMECH Journal 2020)

Cyber-Physical Systems for Construction Sites #

Field robots require sensing systems that remain reliable under vibration, changing terrain, and other constraints of outdoor construction sites. This work develops distributed sensor networks and a ROS2-based cyber-physical platform for collecting and visualizing site information, together with vibration-based methods for assessing ground stiffness during compaction. Work done during postdoc at Kyushu University.

Distributed sensing and VR visualization for earthwork sites.

Sensor Pods and ROS2-TMS (SSRR 2024)

Ground stiffness estimation using ground-mounted accelerometers.

Evaluation of Ground Stiffness (ISARC 2023)

Terrain Understanding for Planetary Rovers #

Terrain understanding is an important component of autonomous navigation for planetary rovers, where visual appearance changes with illumination and thermal sensors may not be available. The work includes models that fuse visible and thermal imagery for illumination-robust terrain classification, as well as methods that estimate thermal information from RGB images, surface geometry, and solar-radiation measurements. Work done during my PhD internship at NASA/Caltech JPL.

Virtual IR sensing for terrain classification and thermal-inertia estimation.

Virtual IR Sensing (RA-L 2020)

Visible–thermal fusion for illumination-robust terrain classification.

TU-Net and TDeepLab (MIPR 2019)

Thermal IR image estimation from RGB imagery.

MU-Net (CVPR Workshops 2019)