RESCUE SIMULATION ANYANG NAVIGATION & MAPPING WALL TOKEN DETECTION ABOUT OUR TEAM Naeun Kim - Team Leader & Lead Developer Nahyun Kim – Navigation & SLAM Engineer Jihee Jung - AI & Computer Vision Engineer Seojun Yang - -Research & Testing Engineer Mingeun Kim - COACH Awards: 2026 Robocup Korea Open 1st in Rescue Simultaion We used Python as our development language, and the total preparation period for the competition was approximately 13 weeks. Autonomous Exploration Strategy Victim Detection Strategy ROBOT INTRODUCTION Environment Representation The robot utilizes a grid-based mapping approach, dividing the environment into 12 cm cells. The robot detects two categories of victims using dual front-angled cameras (Ca1: left, Ca2: right). Trigger Conditions Each cell maintains real-time updates via: Detection is initiated when either camera satisfies one of the following: GPS LiDAR LiDAR-based Cloud Map: Monitors wall occupancy and structural obstacles. YOLO victims (H/S/U): White pixel ratio ≥ 4% in camera image AND corresponding side distance sensor ≤ Color Sensors: Detects floor conditions to identify traversable surfaces. 0.15 m Determining the slam and mapping Status Tracking: Categorizes exploration status into Explored (region entered) Circle victims (C/F/O/P): Concentric circle detected with estimated distance ≤ 0.08 m AND side distance robot's position within Detecting walls for mapping and Visited (center of the cell reached). sensor ≤ 0.08 m our custom virtual cell Each tile is limited to a maximum of 2 detection attempts to prevent redundant processing. distance sensor 1,2,3 Path Planning & Exploration Strategy The system constructs a topological graph by treating cell boundaries as nodes YOLO-Based Detection (H / S / U) and connecting adjacent nodes as edges. Targets with white backgrounds (Harmed, Stable, Unresponsive) are classified via YOLOv5. Front Obstacle detection Frontier Selection: Nodes adjacent to unvisited cells are designated as frontiers for exploration. 1.Advance toward victim wall (670 ms, center-maintaining PD control) Navigation: The A algorithm* is employed to calculate the optimal path to the target frontier. 2.Align robot perpendicular to the wall using LiDAR linear regression on wall point cloud Cost-Aware Pathing: A 8x cost multiplier is applied to edges within swamp regions to discourage traversal 3.Three-angle sampling: capture at 0°, +15°, −30° relative to aligned angle distance sensor 4,5 and encourage path optimization around these areas. 4.Run YOLOv5 inference (confidence threshold: 0.78, max detection distance: 0.25 m) at each angle 5.Majority vote: if the same class appears in ≥ 2 of 3 shots→ confirmed victim; otherwise discarded as false Back Obstacle detection Exception Handling & Safety positive Obstacle Avoidance: If an obstacle is detected at the same coordinate more than three times, the Concentric Circle Detection (C / F / O / P) corresponding node is permanently pruned from the graph. Victims marked with colored concentric rings are classified by ring color analysis. Hazard Response: Upon detecting a hole, the robot triggers an immediate reverse maneuver, marks the area 1.Advance + LiDAR perpendicular alignment (same as above) as impassable, and forbids further re-entry. 2.Fit ellipses to four color masks (Red, Green, Yellow, Blue) and determine outer radius Termination Condition Exploration is concluded once all valid frontiers are exhausted. The robot then 3.Divide the circle into 5 radial bands; sample 180 angles × 8 radii per band 4.Vote dominant color per ring; apply scoring: Red = −1, Yellow = 0, Green = +1, Blue = +2, Black = −2 autonomously returns to the starting tile and transmits a completion signal. 5.Sum of 5 ring scores → classification: 0 = F, 1 = P, 2 = C, 3 = O Self-made Half-Tile Wall Testing Map color sensor ▼CloudMap Floor color detection inertial unit YOLOV5 Direction and angle measurement camera sensor 1,2 Used to align the robot Victim detection to a desired angle ▲Node Graph ▲GridMap Self-made Curved Wall Testing Map Victims are classified using a YOLOv5 object detection model trained on Webots screenshots and Greek letter datasets. Initial testing achieved 87.1% validation accuracy, but only 3 of 15 victims were detected in the simulation environment. To improve performance, unlabeled images were removed, robot POV images were added, Adaptive Equalization replaced Histogram Equalization, and augmentation settings were optimized. After retraining, the model achieved 86.5% validation accuracy but successfully detected 15 of 15 victims in Webots. ▲GridMap