United States of America Pi-rates - RCJ Rescue Maze 2026 Software Architecture Alice Ma (Captain) Software Showrya Verma Romit Gurao Software Hardware Rookie team Developed breath- Developed cognitive Designed robot first search, accumulated error, target and letter chassis, designed Won 1st victim identification dropper assembly, competition as a and obstacle algorithms picked out parts & avoidance team at RoboCup Are there algorithms built robot Junior USA 2026 System setup; mark current unvisited tiles? No Use Breadth-first-Search Navigate to start tile and Check through (BFS) to create a path to initiate exit bonus tile as start BNO055 IMU - Hardware tile attribute start tile sequence Used for ramp detection detection & heading tracking Rescue Kit Yes MG90D Servo Motor - Tilts Dropper Scoring Item Handler & rescue kit Obstacle Avoidance dispenser ramp Rescue Kit Dropper Mechanism: Our robot utilizes a rotary VL53L4CX Wide rescue kit holder that drops kits through a hole when Angle ToF - turned by a stepper motor. We then use a servo motor to Localization Error Obstacle Avoidance Used to detect tilt a ramp, after which the rescue kit slides down to the = 3 obstacles appropriate side. This mechanism was tested extensively tof1 = ( wall on right ? back_tof : front_tof) Colored tile detection for consistent and well-placed dropping. tof2 = ( wall on left ? front_tof : back_tof) while | differnce = the front-tof, back-tof | > tolerance_dist turn (tof1 > tof2) until the difference < tolerance_dist Obstacle Avoidance Robot Issue: IMU was unreliable for navigation and was Obstacle Lines = ToF infrared beam APDS9960 found to drift about 2°/minute - un Localization Error Right sensor < left and center Color Sensor - Also less than 1 tile → obstacle on right of next tile Solution - Self-alignment for error correction: used in colored The robot aligns to straight walls by comparing Issue: Need premature and accurate info on tile detection readings from two Time-of-Flight sensors. Ramp Detection via IMU where obstacle is in order to mark it in file. Oukeda NEMA-8 Adjusts its angle until both sensors report Solution - Comparative localization: Stepper Motor- equal distances, indicating parallel alignment. Robot compares distance in all 3 ToF Turns rescue kit Front wall detection is used to maintain proper Victim Detection sensors tile spacing and prevent collisions. (Computer Vision) All 3 report same dist → wall holder Battery Mount Obstacle can be located based on which sensor returns lowest distance Letter Recognition (YOLO) Cognitive Targets Pololu motors & encoders - Used Ring values: for exact movement -2, -1, -1, 2, 2 Black, Red, Red, Neopixel Light Blue, Blue Ring - provides consistent Battery Mount: Our robot’s main battery Sum = 0 (Stable) lighting for is mounted on its underside, held in Blink LED but cameras Left: Example augmentations place by shelves on our motor mounts. Top: Finished product do not drop kits This placement of the battery makes the The You Only Look Once (YOLO) framework creates predictive bounding VL53L0X (x7) Raspberry Pi boxes based on user-made training images. To improve its accuracy, we We used binary inverse thresholding and HSV values to recognize the Time-Of-Flight- Camera (x2) - robot’s center of gravity lower, added augmentations to the training data, such as brightness changes and cognitive targets and sum up their rings. Binary inverse thresholding Wall / obstacle used for victim improving its capability to traverse zoom. Zooms simulate the robot being closer/further from the letter, and transforms the target’s pixels to white, allowing a contour of the circle detection recognition ramps and stairs without flipping. It also brightness simulates different light levels and shadows. These to be made. Then, the center and radius obtained from those operations makes the robot more compact. augmentations help the trained YOLO model become more robust. are used to extract a pixel from each ring for its value.