Rescue Line U.S.A Team L11 Team Information: Hardware: Software: We are Team FATEN and we will be representing The United States of America in The 2026 FATEN robot is built around one idea: repeatable control Robocup Junior Rescue Line 2026! We are a global team, combining expertise and in unpredictable terrain. It uses a compact 140×180mm chassis with Software overview insights from all over the world, and from multiple previous teams. Therefore we a flexible enhanced-acrylic suspension plate that keeps all four wheels in have a long and wide spanning track record in all types of robotics. We placed 16'th, contact with the ground even when tile height changes. Each wheel is The software runs on three boards in two languages. The XIAO ESP32-S3 runs the camera vision, the K230D runs the AI model, and the Teensy 4.1 is the conductor that decides what the robot does. The Teensy and the XIAO are both Arduino and won the comunity award at the 2024 world championship. We have also independently driven by an STS3032 serial servo. The mechanical (C/C++); the K230D is MicroPython. The two co-processors do the heavy seeing, so the Teensy is left to think and act. seccured three victories at american regionals. structure includes a lightweight dual-grip rescue arm and a Beacuse of this simplified arcitecture, with outsourced tasks, and a desition-making state machine, our main loop front-mounted, linearly guided debris-removal unit. The robot uses becomes really simple. The main loop is only consists of our state machine and its flowchart, explained below. two separate cameras: a downward-facing line-following camera mounted under the front unit, and a front-facing 850nm IR camera for victim detection. The electronics are distributed across custom-designed PCBs and three controllers with distinct State machine Asia Pacific 2021 Asia Pacific 2023 World Championship 2024 roles: a Teensy4.1 for motor control, sensing, and state logic; a XIAO The robot’s behaviour is controlled by a state machine. It spends ESP32-S3 with an OV2640 for line and feature detection; and a K230D running most of the run in its main line-following state, and switches to other states only when certain events occur — for example when a YOLOv8n model for evacuation-zone object detection. Kent Nakai - Head developer it detects an obstacle, loses the line, or enters the evacuation zone. Simple actions like U-turns and green-tile turns are handled This is my tenth year in RoboCup. When I started at nine, I only directly inside the line-following state, since they don’t require wanted to make a robot move forward. Since then, competition has changing to a different mode. Each state is a small, self-contained pushed me to build robots that adapt to their environment and to block of code. This makes the system easy to maintain: adding a understand what it means to build something as a team. After new behaviour or fixing an existing one doesn’t affect the others. moving to the US four years ago, I built teams from scratch at three different schools and won U.S. Nationals three years in a row, though not without many failures. Last year, our robot could not even drive straight through an entire competition. Through all Layered structure of this, what grew the most was not only technical skill, but my understanding of The Teensy code is organised in layers, where each layer mainly other people and myself. This year brings all of that together. With Hampus, whom talks to the one below it. At the top, the state machine decides I met in the Netherlands, friends I used to compete with in Japan, and friends I made what the robot should do. Below that, the actions layer handles the blocking motions such as PID-based driving, turning, and the in the US, we came together to build one robot that balances hardware, software, arm. The processing layer is where raw sensor data becomes and funding. something the robot can trust. It works in four steps. Configuration applies calibration and thresholds so the same Yuna Tamura - Hardware and Management code adapts to different arena lighting. Preprocessing decodes I focus on hardware development, while also handling most of the and normalises the raw bytes from the XIAO and K230D. Filtering removes noise — for example, a command is only accepted after managerial work, as well as taking responsibility for sponsorship a 15-frame majority vote. Finally, fusion combines the cleaned and public relations. Hardware support is one of the roles I’m most values into a pose estimate and a map that the state machine can proud of — I genuinely enjoy building, problem-solving, and use. working hands-on with the team. Beyond the technical work, I care a lot about supporting the team in every way I can. Sometimes that means coordinating logistics, designing our team apparel, or photographing our activities so we can share our progress with Innovative solutions: Custom Power PCB AI integration The main evaluation focus was the evacuation-zone AI model, others. I’ve competed in RoboCupJunior in Japan, and in 2021 I placed 2nd in Rescue We fix last year’s noise in the because slope driving had already been tested through the Line Primary at RCAP Nagoya. That experience shaped how I approach teamwork Debris-remover layout itself. The motor power mechanical layout, suspension, and terrain-biased PID controller. For AI, the difficult part was not only training a detector, but today: combining technical development with the behind-the-scenes work that keeps Our robot features a frontmounted and the logic power sit on their own copper, and both grounds come back to one star point by the battery proving that the model still worked after preprocessing, everything running smoothly. linearly guided debris-pusher quantization, and conversion to a K230D kmodel. Our evaluation capable of following uneven terrain input, so the motor’s current spikes do not move the logic ground. The bulk and pipeline started from raw RGB images collected on the robot. We while still blocking debries from decoupling caps sit right at the regulator and at each chip’s power pin, so the then generated several dataset variants in Python, including raw entering the cone around the line- switching current stays in a small loop and does not radiate into the sensor lines. RGB, shadow-separated images, and CLAHE / histogram-equalized Hampus Mollgård - Mentor and advisor following-camera. This debris- For heat, the regulator pad ties into a large GND copper pour through a grid of thermal vias, so the ground plane spreads the heat out instead of it building on images. Each variant was used to test whether preprocessing remover is based on LWL7B linear improved victim contrast before model training.After that, we With a long background as team leader of a Swedish robotics team, I guide rails and actuated under its own weight and gravity. the chip. With this the supply stays stable and the controller does the same thing every run. trained and compared several detector families, including first met Kent, and Team FATEN as a competing team in Eindhoven 2024. Since then, we have stayed in contact and developed this international collaboration. I reprecented Sweden both in the 2024 Half-duplex IR Camera - world championships finnishing 19'th, and the 2025 world championship finnishing 8'th. We also won the 2025 comunityaward. servo driver Victim detection Thanks to everyone who has supported us throughout this The KRS3301 and STS3032 The 2026 rules introduced are half- duplex serial perpendicular LED lights in the Joshua Enelamah - Advisor servos: each uses one bidirectional data line at 5 V logic, while the Teensy has evacuation zone, so we upgraded to journey — teachers, friends, parents, previous separate RX and TX pins at 3.3 V. We bridge both gaps with a 74HCT126 an IR-based camera. We built it I competed alongside Kent last season, and I saw first-hand how many tristate buffer. When sending, the HCT buffer reads the Teensy’s 3.3 V TX as a ourselves by removing the RGB filter teammembers, competitors, sponsors, and the RoboCup Federation! ideas this team wanted to fit onto a single robot. That ambition exited valid HIGH and drives the line at a clean 5 V, and its tristate enable releases and adding an 850nm bandpass lens. Since LEDs emit almost no IR, the field me to want to keep working with him and FATEN. This year I continue the line so the servo can reply without contention. When receiving, the servo’s lighting has minimal impact on the image, even at competition brightness. In IR, that collaboration from last year mainly on Electronics and Software 5 V output is dropped to 3.3 V through a voltage divider before it reaches the silver reflects strongly while black absorbs, making victim detection much Teensy. One UART now drives both servos cleanly, level-shifted in each cleaner. We first tested 940nm, but it interfered with our ToF sensors. Switching Ideas, so the rest of the team can push those ideas as far as the direction. to 850nm and adding a matching bandpass filter removed that issue. With this hardware will allow. setup, the AI model produced promising results.