Funded research projects I have worked on at GIST AILAB, most recent first. Each entry carries a one-line TL;DR of what I actually built.
Leading the perception stack for a robot that crosses between water and land, where the sensor domain shifts abruptly at the shoreline. Targets: 5 object classes at 65% detection and 90% recognition rate.
Leading the design and collection of a labelled manipulation dataset — trajectories, contacts and object states — for learning-based robot manipulation.
Learning a tactile representation that lets a multi-fingered hand grasp and reorient unseen, irregular objects without a prior CAD model. Targets: 85% peg-in-hole success over 5 object types, 90% object recognition.
Real-time registration of observed RGB-D parts against reference CAD models, driving an AR overlay for on-line industrial part identification and pose estimation.
RGB-IR-LiDAR sensor-fusion object detection from drone views, plus construction of the multi-modal database behind it. Targets: ≥0.85 mAP@0.5 on targets as small as 30×30 px, holding ≥0.65 mAP under image degradation.
On-device segmentation of hair and dust for a domestic robot vacuum, constrained to ≤5M parameters — ≥85% pixel accuracy at ≤10% false-positive rate, trained on a purpose-built 10K-image dataset.
Led a two-step continual test-time adaptation method that keeps a mobile robot’s perception working as the input distribution drifts with the weather — 14% mean corruption error on CIFAR-10-C.
Maritime object detection and situational awareness feeding the navigation stack of a swarm of unmanned surface vehicles.
A ceiling segmentation model and dataset for robust floor-plan generation under heavy occlusion, at 85% layout pixel accuracy — the work behind SMRNet.
Closing the simulation-to-reality gap so that policies trained in simulation transfer to a real collaborative assembly cell.
Deep-learning detection and type classification of corona discharge for automated cable safety inspection — later published at IAS 2023 and awarded the MOTIE Minister’s Award.
Object detection inside low-light underground power tunnels, made robust at the feature level via patch-level cycle consistency rather than by enhancing the image — light enough to run on-board a robot with no comms. Built the cart-mounted multi-camera capture rig and authored the contract report.