NLO Rescue CRO Team Team members Sven Ridzak (left) -Sven leads the team and is responsible for the software that runs inside the evacuation zone, including victim detection, victim handling and exit search. He also designed and built the custom made PCB and worked on the overall hardware Evolution integration. At the very beginning, like everything, our robot started with an idea. Jan Ridzak (right) - Jan is responsible for the line We began work on the 3D design and parts testing in September 2025, following software, including image processing of the with one clear goal: to build a robot that is reliable, fast and capable of line camera, intersection handling and red line solving all elements of the Rescue Line track without manual detection. He also designs all 3D printed parts of the intervention. The first version of the robot was built in December 2025. robot in Fusion 360, including the chassis, wheels and Testing and locating problems on this robot needed to be solved in mounts. various ways, such as mechanical adjustments, component or sensor replacements, or even completely changing the concept. We tested Mentors the first robot at the National RoboCup Croatia 2026 in March, and Juraj Kolarić (middle) and Ivica Kolarić (far left) that was where the evolution of the robot was really visible. We identified major issues such as excessive weight, suboptimal Achievments electronics layout and a PCB that did not fit cleanly inside the chassis. Based on these observations, we designed and built the second robot RoboCup Croatia 2026- 1st place in March and April 2026. The second version focused on lower weight, better balance, a more compact electronics layout and a smaller, redesigned PCB. We also added the Hailo AI accelerator and reformatted our neural network to the Hailo format, which made victim detection run in real time. Today, when we look back, we can proudly say that every challenge and every effort was worth it. Our final robot is not only the result of technological advancement but also of teamwork, creativity and dedication of the entire team. CAD Design To facilitate easy adjustments and ease of repair later in the European RoboCup Austria 2026- 3rd place project, all structural components were chosen to be 3D printed. This method allows precise control over the shape and size of the components. We designed the chassis in Fusion 360 and Tinkercad, and built two versions during testing. The first version was a critical role in defining the layout. During testing, errors in design were identified, particularly parts with thin structures that tended to break easily. These flawed designs were then iteratively redesigned and reprinted throughout the development cycle, ensuring robustness and reliability in the final product. The robot is split into two main sections – the lower section holds the motors, the battery and the line camera with its lighting diodes, while the upper section Innovations holds the PCB, the Raspberry Pi, the wiring and the evacuation Custom debris removing tools zone camera. This layered structure improves accessibility and We expected a lot of debris like shredded papers, chopsticks... makes it possible to change parts quickly when something breaks That is why we developed custom debris removing tools that can during testing. be mounted on the front of the robot to make line detection Lower section Robot in Thinkercad easier and to remove slipping effects caused by the debris. We Connection scheme developed two custom tools, one specialised for chopsticks and the other for any type of sponges or shredded paper. The most important thing was that they still allowed the robot to pass bumps and ramps without any problem. We did that by integrating springs in our tool so it can bend but only for objects glued to the track. Custom silicone-coated wheels To get the best possible grip we designed custom wheels with a 3D printed core and a silicone outer layer. The silicone provides Program flowchart high grip on the field surface, shock absorption when crossing Computer Vision intersections and bumps, better performance on ramps and the We trained a custom neural network option to customize the wheel exactly for our robot. We tested for victim detection on around 2000 several silicone hardness values before picking the one that gave images that we collected ourselves on the best compromise between grip and rolling resistance. This is different Rescue Line evacuation one of the main reasons our second robot now passes ramps zones. The images were labelled in with bumps much better then the first. Roboflow and the model was trained Debris removing tools and silicon wheels on a local GPU using the Ultralytics Main program loop framework. We chose a small model The robot uses a software architecture split between two systems. The architecture so that it can run fast Teensy 4.1 runs Arduino C++ code and handles all low-level control: motor enough on the Raspberry Pi together driving, PID, reading the IMU and sensors and controlling the servos. The with the Hailo accelerator. The neural Raspberry Pi 5 runs Python with OpenCV for image processing and the network detects the live victims, the Hailo AI accelerator for the neural network. The two systems communicate dead victim and the collection (drop) via UART — the Raspberry Pi sends the line position, angle, intersection areas. After positioning itself in the type and victim positions, and the Teensy uses this information together centre of the evacuation zone, the with its own sensor data to decide how to move at every moment. The robot starts spinning slowly and whole program runs in five states. In State 1 the robot follows the line. The searches for victims. The robot first Raspberry Pi reads the line camera, applies a colour threshold to isolate the finds the two live victims and rescues black line, finds its contour and calculates the line centre position and PCB in Easy EDA software them together, and only after that it angle. These are sent to the Teensy, which runs a PID controller that steers searches for the dead victim. Used the robot back to the centre smoothly. A custom diode lighting ring around PCB Design together with the Hailo AI HAT, the the line camera keeps the image consistent in any room. While following the To accommodate various components and model runs in real time while the robot line, the Raspberry Pi also checks for green markers for intersection turns, connectors, we designed a custom PCB in is moving, which makes the system the Teensy monitors the ultrasonic sensor for obstacles and the IMU for EasyEDA. The PCB minimizes the number of more responsive and improves ramps, and automatically adjusts speed and direction when needed. When cables, resulting in a more compact and detection of victims close to walls – the silver line is detected, the robot stops and moves to State 2. In State 2 organized design, while also efficiently the hardest case in our earlier tests. the robot enters the evacuation zone, positions itself in the centre and managing power distribution. Component Progres of the NN model in training starts spinning while the neural network searches for victims. When a placement was carefully planned so that signal victim is found, the robot drives towards it and picks it up with the arm — a routing is short and power and signal lines do photodiode on the arm confirms the victim is captured. Next it enters State not cross unnecessarily. We developed two 3, in which the robot first rescues the two live victims and places them in versions of the PCB. The first version was too the live victim area, then searches for the dead victim (state 2) and drops it large and did not fit optimally inside the in the dead victim area. Once all victims are delivered, the robot enters chassis. As the robot changed shape, the State 4 and uses the IR sensors to follow the wall until it finds the exit and mechanical and hardware requirements leaves the evacuation zone. changed too, and so did the structure and Soldered PCB IMU form of the PCB itself. The second version was The IMU is used for precise turns, orientation and ramp detection. It helps the robot redesigned to be significantly smaller and know when it is climbing a ramp, when it changes angle, and when it needs to recover more efficient, which improved overall from slipping or unstable movement. reliability and made the whole robot lighter. Sensors Lidars JST XH connectors and standard headers are Line CAM The lidars are used to measure distance from walls and objects. They help the robot used for stable connections – the power The line camera is used for detecting the black line, intersections, navigate inside the evacuation zone, detect obstacles and keep the correct distance converters, motor drivers and cameras are green markers and the red line. It works together with our custom while moving. directly mounted on the PCB using specific diode lighting, which keeps the image stable in different rooms andSwitches sockets to interface with the Teensy, the lighting conditions. The switches are used as simple contact sensors and control inputs. They can be used Raspberry Pi and the sensors, ensuring Evacuation Zone CAM to start the robot or detect when a mechanical part has reached a certain position. straightforward and secure connections. The evacuation zone camera is used when the robot enters the Photodiode evacuation zone. It gives the Raspberry Pi a wide view of the zone, so The photodiode is mounted on the victim handling arm. It is used to confirm that a the neural network can detect live victims, dead victims and drop victim has been successfully picked up by the robot. areas.