• Tuton League Region Team Info Sponsors Rescue Maze Japan About Team Tuton is a team of three high school students from Japan competing in RoboCupJunior Rescue Maze. Hardware 1 ×2 M5 2 ×2 Wall Detection Board 5 STM32 F446RE ×4 Twisting Mechanism Floor Color Detection To overcome 2 cm bumps, we designed a compact The color sensor ( 8 ) may falsely detect black tiles 3 2 Tactile Switch UnitV M12 T-mini Plus STM32F446RE VL53L4CX twisting chassis that allows the front and rear sections when the robot is lifted by bumps, while the camera Until the 2025 season, we participated in Rescue Line, where we developed strong skills in reliable hardware design and robust software engineering. Building on this experience, we have taken on the 3 of the robot to move independently. This keeps all ( 9 ) alone cannot reliably distinguish black tiles from challenge of Rescue Maze this year. MAIN Board 5 wheels in contact with the ground and improves silver tiles. To address these limitations, we fused data 4 LED Feetech STS3032 stability on uneven terrain. The mechanism also from both sensors and classified a tile as black only Beyond achieving success in competition, we are committed to contributing to the community. We STS3032 STM32 F446RE ×4 6 minimizes positional deviation while turning on bumps, when they agreed. This approach significantly reduced 6 1 y actively share our knowledge through technical articles and openly publish robot data and development 5 11.1V 1300mAh Z x enabling reliable navigation in challenging maze both false positives and missed detections. results. For more details, please scan the QR codes displayed at the top of this poster. BNO085 STM32F446RE Tower Pro environments. Li-Po Battery MG90D LED MG90D ×2 7 Color Sensor This World Championship will be the final competition for our team. While striving for the championship title, we also hope to connect with competitors from around the world, exchange ideas, and gain valuable Display Board 4 5 experiences. Awards STM32 F446RE 1.54 inch Display STM32F446RE ×4 Raspberry Pi 4 model B TSD10 RoboCupJunior Japan Open 2026 [Maze] • 2nd Place Color Detection Board 8 Reflectance Sensor Camera UART • Best Poster Award 9 9 11.1V • Outstanding Software Award 8 STM32 F103C8 SPI 7 I²C RoboCupJunior Japan Open 2025 [Line] S9706 STM32F103C8 LBR127HLD Raspberry Pi Camera Module V3 • 2nd Place Deformable Wheels Rescue Kit Deployment • Outstanding Hardware Award Japan Open 2025 Japan Open 2026 To improve stability on bumps, ramps, and stairs, we The rescue kit is an 11 mm cubic capsule manufactured 1 Bumper 5 PCB Unit 6 Wheel ×4 Members Two tactile switches are mounted on the left and right The main electronics module containing the robot's The wheels are driven by STS3032 serial servo motors. developed a wheel ( 6 ) that provides suspension using a 3D printer. A freely moving nut is enclosed sides of the robot. They are used to detect collisions primary sensors, processors, and communication Each wheel is a custom-designed deformable wheel through deformation of the wheel itself. Unlike inside the cube to absorb impact energy when the kit with walls and obstacles. interfaces. featuring a silicone-molded tread, providing high conventional spring-based suspension systems, this contacts the floor or walls, reducing bounce and approach requires little additional space. The 80 mm helping it stay near the target location. Shuji Igarashi traction and excellent terrain traversal capability. A. Raspberry Pi 4 model B The main processor Captain, PCB, Software 2 UnitV Camera ×2 wheel is constructed from TPU, PLA, and silicone. Its The robot carries a total of eight rescue kits. Four kits Two UnitV cameras are used for victim detection. LEDs responsible for executing the robot's core software, 7 Rescue Kit Releaser ×2 deformable structure absorbs shocks and maintains including mapping, navigation, and decision-making. The rescue kit deployment mechanism. Four rescue kits are stacked vertically on each side of the robot and an I have been involved in RCJ since 2023. As Captain, I led project management and provide consistent illumination independent of ambient traction on uneven terrain, enabling reliable traversal of MG90D servo motor pushes the kits out one at a time, system integration. I also designed and manufactured our custom PCBs and lighting. B. T-mini Plus A 360-degree LiDAR sensor used when are stored vertically on each side of the robot and are all obstacles encountered in the Rescue Maze field. released individually using servo motors. enabling controlled and reliable deployment. contributed to the software, including the exploration algorithm. wide-range environmental measurements are required, 3 Top LED such as distinguishing walls from obstacles and 8 Color Detection Board Load A status indicator LED that flashes when a victim is generating local surroundings data. Silicone Ryouta Koudu Used for tile classification. Visible-light sensors are used detected and when the robot successfully returns to the Hardware start tile. C. VL53L4CX A single-point time-of-flight distance to identify black, blue, and red tiles, while infrared reflectance measurements are used to detect silver tiles. TPU sensor used for high-speed and precise wall distance I have been in RCJ since 2023. As the Hardware Chief, I designed the robot's hardware measurements. 4 LCD Display 9 Front Camera using CAD and was responsible for its mechanical assembly. In particular, I focused on PLA developing a robust drivetrain capable of navigating the challenging terrain of the field. Displays the robot's current operating status and D. BNO085 An IMU sensor used to estimate the An auxiliary camera used as a supplement to the Color position within the generated map. robot's orientation and heading. Detection Board to assist in color tile identification. Shuichiro Mitsumoto Software, Camera I have been participating in RCJ since 2025. As Software Chief, I was responsible for A Exploration Algorithm F 3D Mapping software architecture and implementation. In particular, I focused on image processing to improve victim detection and support Cognitive Targets introduced this year. Software The main control logic is implemented in Python and executed on the Raspberry Pi. The flowchart on the left shows the overall software The robot repeatedly moves to the nearest unexplored tile using the To accurately map multi-level environments, we developed a LoP Handling flow. lowest-cost path and returns to the start tile after completing height-based layer matching system. New layers are generated Development Process Start LoP Monitoring (Active During All Operation) G Detect LoP The robot explores the maze one tile at a time while continuously updating its internal map based on exploration. Path planning is based on Dijkstra's algorithm, with a X 90° Y dynamically whenever the robot discovers a new floor through a ramp. Vertical displacement is estimated using gyro data and traveled distance, allowing layers at similar heights to be matched With less than one year to transition from Rescue Line to Rescue Maze, our primary challenge was to detected walls. When a checkpoint is reached ( E ), the graph node representing both position and orientation. and treated as a single layer. 0° fully understand the competition requirements and develop a robot capable of meeting them all. To Wait for Restart current map and exploration state are saved. In the By incorporating turning costs as well as movement Detect Surrounding Walls event of a Lack of Progress (LoP, G ), the robot costs, the planner more This approach enables reliable mapping even in complex fields achieve this, we adopted an iterative development approach, rapidly identifying issues, testing potential 90° with multiple interconnected ramps of varying inclinations. To solutions, and refining both the hardware and software through continuous experimentation and Estimate Current Heading restores the most recent checkpoint data and resumes accurately captures the X Y Update Map operation. During each tile transition, the robot robot's real behavior. 180° prevent mapping errors caused by stairs, the system evaluation. distinguishes stairs from actual floor transitions and avoids Restore Checkpoint continuously performs black tile detection ( B ), To accelerate development, we built two robots, allowing hardware and software development to Search for Unexplored Tiles obstacle detection ( C ) and victim detection ( D ), Simulation experiments on randomly creating unnecessary layers when no layer change has occurred. proceed in parallel. Team members independently explored solutions to challenges, enabling fast executing the appropriate response whenever one of generated maze fields showed that Layer 2: Altitude 30cm iteration. Key software components, including victim detection and autonomous exploration algorithms, these elements is identified. our method explores more were also developed and tested on laptops before hardware completion, reducing later integration work. efficiently and consistently than the YES NO Extended Right-Hand Rule, Unexplored Tile After all reachable tiles have been explored, the robot Exists? Shuji Exploration PCB #1 PCB #2 PCB #3 Practice on Field calculates the shortest path to the start tile and achieving lower overall exploration returns to complete the run. costs. Ryouta Hardware #1 Hardware #2 Hardware #3 Hardware #4 New Wheel A Calculate Path to It Shuichiro Camera (2025 Victims) Integrate Practice on Field Camera (2026 Victims) A Calculate Path to Start The source code is available here: A demonstration of the exploration algorithm running in Japan Open World Championship Heading Correction the simulation environment can be viewed here: Layer 1: Altitude 0cm For hardware development, we used Autodesk Fusion for mechanical design and KiCad for PCB design. Return to Start Tile by Tile Sharing all design data within the team enabled efficient task distribution and collaboration. LED Blink and Finish D Victim Detection For software development, we used Visual Studio Code and GitHub. The Raspberry Pi software and Forward Motion UnitV software were written in Python, while the STM32 firmware was developed in Platform IO using Return Home 1. Letter victims 2. Cognitive Target the Arduino framework. To maintain code quality and teamwork efficiency, we emphasized readability, B Detect & Avoid Black Tiles documentation, and a well-structured codebase. Captured images are binarized and converted to Candidates for cognitive targets are first detected from the image using a flood-fill C Detect & Avoid Obstacles grayscale before being processed by a MobileNetV1 method. Shape-based filtering is then applied, and the largest valid candidate is HARDWARE SOFTWARE (α=0.75) model. A victim is detected only when the selected. After estimating the radius and center of the selected candidate, 3x3 D Detect & Rescue Victims model output exceeds a threshold across consecutive pixels square regions are sampled in 12 directions, centered at points whose frames, reducing false detections. distances from the center are 0.1, 0.3, 0.5, 0.7, and 0.9 times the radius. The color To improve robustness, data augmentation techniques of each layer and the total score are determined by voting based on these samples. NO Reached Next Tile? such as CutMix and image flipping were applied during After that, the target is considered detected only when the agreement level of training. In addition, Out-of-Distribution (OoD) these votes exceeds a threshold and the same score is detected across consecutive Fusion KiCAD GitHub VSCode PlatformIO Python C++ YES Tile Transition detection using the Maximum Logit Score rejects low- frames. confidence predictions, improving reliability in real- Throughout development, we documented issues and action items in our Engineering Journal, allowing us world environments. Experimental results showed that Why we adopted a rule-based approach rather than CNN-based methods: E Save Checkpoint to clearly track our progress and maintain up-to-date understanding of the project's status. In addition, this threshold avoided approximately 40% (73/185) of • Running multiple CNN models simultaneously is impractical due to memory and computational constraints. we incorporated the official RoboCupJunior Rescue CMS into our testing and practice sessions, which false detections on fake victims. • Each cognitive target label exhibits a wide variety of appearances, making robust classification difficult. F Update Map Recognition Results helped us analyze the robot's weaknesses and identify areas for improvement. Blue Tile Handling Detect Surrounding Walls • A rule-based approach allows for easier tuning and adaptation compared to a CNN-based approach.