M.S. Student, Mechanical Engineering
Korea University
Humanoid Robotics · Whole-Body Control

Overview

Developed an obstacle detection and avoidance pipeline for the DOGZILLA quadruped robot using monocular ORB-SLAM for depth estimation and UV-disparity maps for obstacle segmentation. The system enabled real-time autonomous navigation in unstructured environments.

Period: Mar.2024 – Jan.2025
Role: Team Leader
Team: RoboStride
Award: President’s Award, Creative Challenger Program (CCP), Korea University — Jan.2025

Problem

Quadruped robots operating in real-world environments face unstructured terrain and unexpected obstacles. Deploying reliable obstacle avoidance without dedicated depth sensors (e.g., LiDAR, stereo cameras) is a key challenge for cost-effective and lightweight robot platforms.

Approach

Depth Estimation via ORB-SLAM

  • Applied monocular ORB-SLAM to estimate scene depth from a single camera feed
  • Extracted sparse 3D point clouds for obstacle localization

Obstacle Segmentation via UV-Disparity

  • Constructed U-disparity and V-disparity maps from the estimated depth
  • Segmented obstacles based on disparity peaks, distinguishing ground plane from objects

Avoidance Behavior

  • Integrated detection output into the robot’s navigation controller
  • Triggered avoidance maneuvers in real time upon obstacle detection

Robot Platform

DOGZILLA — quadruped robot platform
Camera: monocular RGB camera

Skills

C++ Python ROS2 ORB-SLAM OpenCV