Method, Production & Design Software Robot Construction Navigation Algorithm Teensy 4.1: OpenMV AE3: The core architecture of the software relies on a Finite State Machine (FSM) to manage all The main controller that The robot holds two robot movements. An FSM is a system where each task a robot needs to complete is give does all the processing for OpenMV AE3 cameras, a state. When the robot then needs to complete that task, it will then move into that state. the robot is a teeny 4.1. one on the left and one on To then move out of this state there is then a set of conditions that need to be met for it to This was chosen due to its the right. These cameras are change into a different state. At the center of our FSM is the mapping state; as the robot superior speeds compared to used for victim detection traverses each tile, it executes the mapping algorithm to determine its next destination other microcontrollers and and were chosen for the based on available empty tiles. This FSM is also in charge of the process for when a victim its small form factor. small form factor. is detected and along with what task the robot as to perform on tile such as a blue tile. MG996R Continuous Servo Adafruit VL53L4CD: Motor: The robot holds eight The robot holds four VL53L4CD distance sensors, To ensure precise timing, sensor data is captured using hardware interrupts. The main MG996R continuous servo two on each side. These are software cycle runs every 100 milliseconds, during which different sensors are sampled at motors to drive the wheels. used for sensing walls in all varying frequencies depending on their individual reading acquisition times. With different These were chosen due directions and correcting the sensors being read every 5ms with the over 100ms sensor reading cycle. to their ease of use and robots’ skew while driving relatively small form factor. around the maze. Adafruit VL53L4CX: 180-degree 9g Servo Motor: DELTA The robot holds two The robot uses three 9g VL53L4CD distance sensors servo motors, two for the for long distance sensing in LRF pivot and one for the large open spaces of mazes. package delivery system. Our mapping strategy combines a standard left-wall-following approach with a custom Limit Switches: Adafruit AS7341 colour navigation algorithm. The robot follows the left wall while traversing unvisited tiles. The robot uses four limit sensor: However, upon reaching a previously visited tile, the custom algorithm takes over to switches, two on the front The robot holds one colour determine the next move. This algorithm checks if any adjacent tiles connected to The structural design and PCBs of the robot were created in and two on the rear. These sensor on the bottom of the the current visited tile remain unvisited. If an unvisited neighbour is found, the robot Autodesk Fusion. Using Autodesk Fusion allowed for the robot to are used for centering front for detecting colour navigates to it and resumes the left-wall-following algorithm. If all adjacent tiles have be visualised at any stage in the design before manufacturing it. The the robot and detecting floors such as red, black already been visited, the robot backtracks to the previous tile and repeats the check. This robot was designed and manufactured completely in-house with the obstacles. exception of the PCBs, which were designed in-house but sent to JLC PCB to manufacture. The PCBs were still hand-soldered in-house. GY-BNO055 9 axis IMU: The robot uses one BNO055 or blue. This sensor was chosen based on its large Brisbane BoysCollege - Rescue Maze backtracking loop continues until an unexplored path is found and eventually the robot returns to the start tile. Throughout this process, crucial data such as tile coordinates, range of colour channels (10 coloured tiles (e.g. blue or black tiles), and connectivity to other tiles. This robot used many manufacturing techniques, including 3D IMU for detecting the channels). printing on a Bambu Lab P1S and CNC milling on a Makera rotation of the robot in three Carvera. axis. The current robot uses nine major components, these can be seen on To optimize memory usage, the robot implements dynamic memory allocation for the right. mapping instead of using a 3D array or static list. This approach ensures the robot only consumes the exact amount of memory required, preventing waste. Furthermore, utilizing dynamic allocation shifts the data storage from the stack to the heap. This is crucial for the Teensy 4.1, which limits stack size to 16–32 KB while providing a much larger ~512 KB heap. Within this heap space, each map tile is represented as a structured PCBs data type (struct). Each tile struct stores its unique ID, coordinates, colour data (e.g., blue, black, silver), and victim presence, alongside pointers to adjacent tile structs to form the map network. The main PCB of the robot was designed completely in Autodesk Fusion because of the simplicity of Victim Identification designing every hardware part in the same program and the ability to visualise the 3D virtual PCB with the robot assembly. This PCB was made to connect every signal from sensors and motors to the teensy 4.1 microcontroller. This board also houses the power circuit which consists of the 3.3V and 5V regulators. The entire power circuit is located on the back of the board to allow for easy testing and fault nding when an issue arises, especially when the issue is a short from power to ground. An image of the main PCBs layout can be seen on the le­. The Light PCB of the robot was also Abstract To detect the letter victims we are using a FOMO MobileNetV2 object detection neural-network for both the letter victims and the cognitive targets. This designed completely in Autodesk Fusion DELTA is an Australian team from Brisbane Boys’ College competing in Robocup Junior Rescue Maze. We are a team of three members and have competed approach was chosen after seeing its capabilities in 2025 for the same reasons as the main PCB. in the Rescue Maze category for two years. Development of the current robot started in late 2025 where every team member has put in over 300 hours each at internationals where it was observed that many teams The robot holds two of these light PCBs, to design, build, program and test our robot. DELTA competed in the Robocup Junior Rescue Maze International competition in Salvador, Brazil in 2025. that used a similar approach achieved very good results one on the le­ and one on the right. This We have since competed in and achieved first place in the Robocup Junior Open Rescue Maze National Competition in 2025 and the Robocup Junior Open and stood out for their victim identification. This year board lights the camera shot and allows Rescue Maze Brisbane Regional Competition in 2026. over 500 images were taken on an OpenMV AE3 camera for the cameras to observe the victims The robot was made using industry standard technologies including - 3D printing, CNC Milling and laser cutting, along with the use of printed circuit of all the possible letter victims and these were used to in almost exactly the same lighting boards (PCBs) to improve reliability and save time when constructing spare boards. train a model inside of Edge Impulse. Using the OpenMV conditions no matter where the maze is Our team is made up of three members, Ethan Seymour, James Cousins, and Connor Duncan. AE3 camera helped replicate what the victims would or how bright the lighting is. This board Ethan works on the hardware, including electrical design, structural design and manufacturing. Ethan has designed every part of the current robot himself look like on the robot by using the exact same hardware includes a 15 individual white LEDs and has manufactured most parts himself with a CNC machine or 3D printer. James works on the main software and navigation algorithm and has coded as mounted on the robot (same camera and lens). These and a potentiometer to ne tune the every part of the teensy code himself. Connor works on the victim detection software and has trained and coded the cognitive target recognition himself. images were then trained in Edge Impulse with 50 brightness of these LEDs. An image of epochs, a learning rate of 0.003 and a validation set size of the light PCBs layout can be seen on the 20%. This was tuned from the default values to achieve a right. higher precision score, recall score and therefore a higher F1 score. The image on the left shows the generated image features. Pivoting LRFs The target detection works in two steps, it first uses blob detection to locate a target by looking for a large area of colour. Then it uses ring detection code, which works in two main stages, to identify and interpret the target. The camera searches for the target area, locks onto the outer circle, and keeps that circle position stable across frames so the readings do not jump around from one frame to the next. After the target has been The pivoting LRF system was implemented to allow the robot to always see the wall straight on even when located, the code reads the five concentric rings by sampling several fixed points around each ring in the form of dots and checking which colour the robot is tilted up or down. This system utilises a single 180-degree 9g servo motor. Image of team DELTA at the Robocup Image of team DELTA at the Robocup Scan the QR code to visit our website threshold each point matches. It then uses a voting method so that each ring gets a final colour decision based on the majority of samples. Once This runs constantly throughout the robot’s running and prevents the robot from misidentifying where it is Junior National competition in 2025 International competition in 2025 for more information. all five rings have been classified, the script converts those colours into their assigned numbered values, adds them together, and uses the total to located because it saw a wall at an angle. decide whether the target is real and if it should drop a package or not. The code also checks that the same ring pattern is seen repeatedly before The robot uses two of these with one on the front and one on the back. Although using a similar system confirming the result, which helps reduce false detections. may be useful on the side facing LRFs, space constraints did not allow for this on the current robot design. Custom Silicone Wheels An image of this software detecting a target can be seen below The robot runs four of the same custom molded silicone Wheel Pivot The Motor Pivot can be seen in the image above. This system consists of 2 needle roller bearings, one ball bearing wheels. These wheels are made from a central 3D printed hub made from PLA and Shore 25A silicone for the tire. Silicone was chosen because of its superior grip due to its elasticity. An added benefit of the silicone wheels is that and an M8 bolt through the center. This bolt is fixed on one side and placed in a bearing on the other to allow it unlike some other choices, silicone does not stick to most to pivot with minimal friction. The other two bearings are placed in the middle to allow the spacer to rotate. things and therefore will not pick up debris while navigating The major reason for the use of a pivot was because of its ability to keep all wheels on the ground at the same around the maze. The chosen silicone has a hardness of time. When a robot has fixed motors, scenarios arise where the robot cannot place all four wheels on the ground 25A on the Short Hardness Scale and was chosen for its and will instead rock back and forth between two opposite wheels. With a motor pivot it is possible for the durability and grip. Other options were considered but were motors to pivot down in these scenarios and therefore keep all four wheels in contact with the ground at all found to be too easily broken or not grippy enough. times. The tire incorporates six large notches that allow for the This design was chosen based on its high strength from the use of the M8 bolt through the center. When DELTA robot to climb stairs. The groves of the stairs catches on the competed in 2025, it was observed that many teams had smaller pivots that at times would flex and would notches and gives the robot additional grip while climbing struggle in scenarios where the robot encounters large amounts of debris. stairs. Images of these wheels can be seen on the left and right.