</> Robotics • Perception → Control

Welcome to my corner of the web.

I’m Micheale Haileslassie Gebrezgiher, an undergraduate student at Chonnam National University (GKS-U scholar) and a research assistant at AIRLab.

desk photo

Selected works

2023 – 2025
Ball Balancing cover
Ball Balancing project
 Stereo DLO
Stereo vision–based DLO tracking
 Splines
Experimental comparison for splines
 UKF DLO
Unscented Kalman filter for DLO tracking
 Dual arm
Dual-arm DLO shape control

UR5e Ball Balancing (PPO)

Ball Balancing cover

Summary

Presented at ICCAS 2025, this project tackled the challenge of real-time ball balancing—a classic benchmark for dynamic robot control. Prior work typically depended on 7-DoF torque-controlled robots and expensive sensing, but we showed it can be achieved on a standard 6-DoF UR5e with a low-cost touch panel.

  • Robot: UR5e, velocity control (PyBullet + real deployment)
  • Sensing: Resistive touch panel for continuous ball position feedback
  • Method: Dataset-guided PPO trained on spline-based trajectories
  • Result: ~87% real-world success, <4 cm average deviation
Environment/Hardware
Environment & hardware overview
Pipeline
Dataset → RL → Simulation → Real robot

Deformable Linear Object (DLO) Manipulation & Tracking

DLO overview cover

Summary

Much of my research centers on deformable linear objects (DLOs), which are challenging to model, track, and manipulate due to their infinite degrees of freedom. My work spans perception, estimation, and control, combining stereo-vision pipelines, spline interpolation, UKF state estimation, and deep learning models (Bi-LSTM, GRU, Transformer).

  • Tracking: Stereo vision + cubic splines + UKF under occlusions
  • Modeling: Bi-LSTM/GRU vs Transformer for DLO dynamics
  • Control: Dual-arm setup for real-time shape regulation
  • Publications: ICROS 2024, ICCAS 2024, KSME 2025, IFAC 2025

Stereo Vision–based DLO Tracking

Stereo DLO setup
3D cable shape estimation flowchart

Summary

Presented at ICROS 2024, this project introduced a stereo vision–based pipeline for real-time 3D tracking of deformable linear objects (DLOs). By combining 2D preprocessing, node synchronization, and 3D reconstruction, the system achieved continuous and robust tracking of cable shapes under manipulation.

  • Setup: Dual Basler stereo cameras + dual robot manipulators
  • Method: HSV preprocessing → node sync → 3D spline reconstruction
  • Performance: Mean error ~0.93 mm (X), 1.1 mm (Y), 4.2 mm (Z)
  • Result: Effective real-time tracking of 7 markers on DLO
Stereo performance
Tracking accuracy in XY/XZ planes
Experiment setup
Experimental setup for tracking Evaluation
Flowchart
Experiment setup and result

Experimental Spline Comparison

Spline comparison cover

Summary

Comparative study of spline-based representations for DLO tracking. Evaluated marker-based data against cubic/B-splines for accuracy, robustness, and runtime in both synthetic and real setups.

  • Method: Cubic vs B-splines for shape interpolation
  • Setup: Stereo vision pipeline with synthetic + real markers
  • Result: B-splines reduced error by ~20% under occlusions
Case 1 result
Case 1: Cubic vs B-spline under mild occlusion
Case 2 result
Case 2: Error profile comparison across frames

UKF for DLO Tracking under Occlusion

UKF DLO cover

Summary

This project addresses the challenge of real-time 3D state estimation of deformable linear objects (DLOs), such as cables and hoses, which are difficult to track due to their nonlinear dynamics and frequent occlusions. The framework integrates stereo-vision tracking, spline interpolation, and an Unscented Kalman Filter (UKF) to achieve robust and accurate tracking even when parts of the object are missing from view. A data-driven Bi-LSTM dynamic model is combined with UKF to fuse predictions and measurements, resulting in smooth, reliable state estimation suitable for robotic manipulation tasks.

  • Real-time stereo reconstruction with dual cameras
  • Spline interpolation to recover missing nodes under occlusion
  • Bi-LSTM dynamics + UKF for robust nonlinear state estimation
  • Validated on 169 trajectories, achieving <5 mm error
Occlusion handling
Handling occlusions
Out of frame
Recovery from out-of-frame

Dual-Arm DLO Shape Control

Dual arm cover

Summary

Real-world dual-arm manipulation of cables with shape regulation. Combined time-series predictions with model predictive control for closed-loop execution.

  • Setup: Dual UR5e arms + DLO testbed
  • Method: Bi-LSTM for state prediction + MPC control
  • Result: ~92% shape conformity within 3 cm deviation
Case 1 result
Experimental environment and software framework.

Deep Learning for DLO Dynamics

Deep learning DLO cover

Summary

Time-series deep learning models (Bi-LSTM, GRU, Transformer) for predicting DLO dynamics under robot actuation.

  • Dataset: 350k+ samples of UR5e cable manipulation
  • Models: Bi-LSTM, Bi-GRU, Transformer
  • Result: Bi-GRU achieved lowest error (~2.6 mm short horizon)

Projects

UR5e Ball Balancing (PPO) Sim-to-real controller using dataset-guided RL
2025
Deformable Linear Object (DLO) Manipulation & Tracking Perception → Modeling → Control pipeline
2024–2025
Stereo vision–based DLO tracking DLo / realtime/ state estimation
2024–2025
Experimental Spline Comparison Marker-based tracking methods
2025
Dual-Arm DLO Shape Control Time-series data-driven control
2024-2025
UKF for DLO Tracking under Occlusion Stereo vision + splines + UKF
2025
Deepfake Image/Video Detection CNN; UADFV & Celeb-DF
2024
Yakgwa Cookie Classification Classical ML; 96% accuracy
2024

Resume

View CV Download CV
Jun 2023 – Present
Education B.Sc. Mechanical Engineering, CNU (GPA 4.18/4.5)
2022–2026
Awards ICROS 2024 Undergraduate Paper Award; SE, Challenge, Energy+ AI scholarships; GKS
2021–2024
Skills Python, C/C++, MATLAB; PyTorch, TensorFlow; ROS2, PyBullet, CoppeliaSim; YOLOv8
Publications (selected) IFAC 2025 (1st), ICCAS 2025 (1st, accepted), KSME 2025 (1st), ICROS 2024 (1st), ICCAS 2024 (co-author)
Languages English (Full professional), Korean (TOPIK 5), Amharic, Tigrinya

Insights & Notes

Ideas, breakdowns, behind-the-scenes.

About Me

I’m Micheale Haileslassie Gebrezgiher, a fourth-year Mechanical Engineering student at Chonnam National University, supported by the Global Korea Scholarship (Undergraduate).

As a research assistant at the Autonomous Intelligent Robotics Lab (AIRLab), my work bridges robot perception and control. I’ve developed real-time stereo vision pipelines, spline-based state estimation with UKF, and reinforcement-learning controllers for tasks such as UR5e ball balancing and deformable linear object (DLO) manipulation.

I’ve authored and presented multiple papers at international venues including IFAC, ICCAS, and KSME. Beyond research, I’ve participated in innovation/startup programs in Silicon Valley, and I’m passionate about applying robotics and AI to real-world problems.

My long-term goal is to pursue graduate studies (MSc → PhD) and contribute to advancing collaborative robotics, human–robot interaction, and intelligent control systems.

about1
about2
about3

Let's create something awesome together.