1
Community Award
TOKAI Ghost
Line league · Japan ·
RoboCup 2026
· 1st place
Document register
- Poster1 pagePublished
- Presentation videoYouTubePublished
- Bill of materials3 pagesPublished
- Team description paperNot shared
- Engineering journalNot shared
- Source code24.7 MB · GitHubPublished
Sharing each document is the team's decision. “Not shared” means this team chose not to publish it, or did not submit one — not that it is missing from the archive.
In their words
We are TOKAI Ghost, a team from Japan. Our new robot for this competition is a rescue robot designed to consistently achieve high scores in all situations. At Japan Open 2026, we achieved full marks in every round.
This robot is entirely self-built. Unlike our robot from last season, we didn’t use any kits such as LEGO for this robot. We also designed all the electronic components.
Our robot has many innovative features. We developed silicone wheels and a twisting mechanism to navigate difficult ramps and bumps. For line following, we use not only sensors but also a camera this year, enabling more stable tracking. In the evacuation zone, the robot can reliably pick up and efficiently carry victims with its multifunctional rescue mechanism.
For victim detection, we trained a YOLOv8-based model on a dataset annotated using Label Studio. Since it was designed for inference on a Raspberry Pi 5, we were able to develop a highly accurate model.
By combining the strengths of a camera and multiple sensors, the robot improves both speed and stability in line following and in rescue tasks.
Poster
Read the text of this document — 1109 words
Hardware
Multi-Processor Architecture
The robot uses a Teensy 4.0 as the main controller, a Raspberry Pi 5 for image
processing and AI inference, and two XIAO RP2040 boards as sub-controllers. The
Raspberry Pi processes camera data, while the RP2040 boards handle dedicated
sensor and actuator tasks. This distributed architecture enables high-speed
control, accurate vision-based recognition, and reliable parallel processing.
Powerful Power Train
The robot is driven by four STS3032 serial servo
motors. These compact yet powerful motors enable
the robot to overcome bumps reliably.
The wheels feature a large diameter, narrow width,
League : Rescue Line
and custom-molded silicone tires. This design Country : Japan
provides high grip on bumps and ramps, improving
traction and driving performance.
Multifunctional OUR TEAM
Twisting Mechanism Rescue Mechanism Our team, TOK AI Ghost, consists of students
The front and rear halves of the robot are Our rescue mechanism can collect, identif y, from the s ame high school . We began
connected by linear shafts, allowing them to classify, store, and evacuate victims within a participating in RoboCup Junior Rescue Line last
twist relative to each other. This mechanism compac t system. B y stor ing all vic tims season. Building on the experience we gained at
keeps all wheels in contact with the floor on simultaneously, the robot reduces travel time and RoboCup 2025 S alvador, we developed a
uneven terrain, improving driving performance improves rescue efficiency. The rescue arm uses powerful, compact, and reliable robot.
on bumps and ramps. It also maintains a stable sensors to detect obstacles, confirm victim
distance between the line sensors and the floor, capture, and identif y victim conductivity. A TAKUMI TAKANO
enabling smooth and reliable line following even
at transitions between flat tiles and ramps.
one-sided gripper and tilting gate mechanism
enable victim sorting and selective evacuation.
Overview Captain, Hardware and Software Developer
We designed the entire robot
Takumi Takano designed all the electronic components and most
structure using Autodesk Fusion and
of the robot’s structure, and he manages the development
manufactured most components using 3D printing to maximize
Rotating Camera The camera is mounted on a servo-driven mechanism and is
primarily used for victim detection. In gap situations, the camera space efficiency. The robot measures 163 mm × 174 mm. An
schedule. He also coded the algorithms for line following and
rescue. Additionally, he coordinates the opinions of team
Mechanism rotates downward and is used for camera-assisted gap control. aluminum chassis provides durability and lowers the center of members.
gravity, while lightweight 3D-printed components are used in the
upper half of the robot. All custom circuit boards were designed HONGYI DAI
Software using KiCad. The robot software was developed using the Arduino
framework, MicroPython, OpenCV, and Ultralytics.
Software Developer
Hongyi Dai coded the algorithms for detecting evacuation
points. He also developed a camera-based assist system for
Line Following Evacuation Zone line following. Furthermore, he is responsible for setting up a
Raspberry Pi 5 environment and configuring the camera.
The robot calculates the line position from line sensor data and follows the line An overview of the evacuation zone process is shown in the figure below. Throughout
using PD control. When it detects an intersection with the color sensor, it the rescue operation, the robot utilizes its multifunctional rescue mechanism and SHINICHI KADO
determines the appropriate turn based on the elapsed time since the last black distance sensors to handle challenging environments with many obstacles. Because
Software Developer
line was detected. On ramps, it applies slope-specific compensation to the robot collects victims regardless of their type, it can perform rescue operations
Shinichi Kado joined the team this season. He developed a
optimize line-following performance for each incline. When no line is detected, efficiently without prioritizing specific victims. The number of rescued victims is stored machine learning model for victim detection. He also developed
it first moves backward and determines whether the situation is a gap or a line on the Raspberry Pi throughout a run. After a LoP, the robot refers to this information,
following error, significantly reducing the risk of LoP. When encountering an allowing it to perform optimally even on subsequent attempts. The algorithm is
Victim Detection efficient methods for capturing images and augmenting data for
the dataset. Thanks to his work in introducing new technology,
obstacle, it uses distance and touch sensors to navigate around it while designed to handle various situations, such as failing to collect all victims or being We previously used an OpenCV-based algorithm for victim detection. However, the robot can reliably detect victims.
maintaining the shortest possible clearance. unable to locate a corner. distinguishing victims from the background was difficult, so we adopted
YOLOv8, a deep learning-based object detection model. Using transfer
learning based on the pretrained YOLOv8n model, the robot can detect both
the position and type of victims from camera images. RoboCup Junior
We collected and processed our own dataset of approximately 3,000 images
and manually annotated all victims. To improve robustness, the dataset JapanOpen 2026
included dummy balls, unlabeled images, and Blender-generated images.
・1st Place ・Software Award ・IRS Award
Additional datasets containing flashlight interference were also created to
address the new competition rules.
RoboCup 2025 Salvador
・Rescue Line 5th Place
Innovation
Although the victim detection model was trained using images containing ・Outstanding Design Award
flashlight interference, false detections still occurred under certain light
Camera-Assisted Gap Control
intensities and angles. FLL FIRST 2025 Championship
To address this issue, we installed a polarizing film in front of the camera lens.
Line sensors provide stable and low-noise line following. However, because This modification significantly reduced strong reflections from the floor and ・Engineering Excellence Award Winner
they can only detect the presence of a line, they cannot determine the improved detection reliability under challenging lighting conditions. ・Encore Celebration Alliance FInalist
direction of the line while traversing a gap.
To address this limitation, the robot uses a camera during gap traversal. When
a gap is detected, a servo motor rotates the camera downward, and the line
angle is calculated using OpenCV's PCACompute function. The robot then Corner Detection Supported by
uses this angle to align itself with the recovery line. The robot detec t s the evacuation point using
By combining sensor-based line following with camera-based directional OpenCV-based image processing. Red and green
analysis, the robot can navigate challenging gaps more reliably. regions are extracted from the camera image, and
connected-component labeling is used to remove
Hi!
noise. The evacuation point is then identified based on X note YouTube
the aspect ratio of the detected region. To improve
robustness when obstacles partially occlude the
evacuation point, the robot evaluates the two largest
labeled regions instead of only the largest one.
1 page, rendered as images so they load quickly. The text above is the document's own, extracted from the PDF.
Presentation video
Hosted on YouTube. The player loads only when you press play.
Bill of materials
Read the text of this document — 902 words
RESCUE JUNIOR - Bill of Materials (BOM)
Team name: TOKAI Ghost
Every Line/Maze team has to submit a bill of materials for their robot
Instructions:
*Enter Team name above
*All costs need to be in local currency and their approximate
conversion into dollars.
*All components worth less can be summarized in one line quantity 1
and their total cost (e.g. Screws)
*In the Local Currency Name column, the team must enter the name
of the local currency, if it is different from US dollars.
*In the software sheet, in the Author column, the name of the
designer, writer, or company that owns the software or library should
be entered. If the team created their own algorithm, dataset, or AI
model, they can include their names as authors.
*On the hardware sheet, in the "Author" column, enter the name of the
company, engineer, or part manufacturer. If the team made a custom
part, such as the robot housing, the team should include their names
as the authors.
*IMPORTANT* The "Hardware" sheet is not required for simulation
Name of the Total Cost Total Cost ($)
Team name: TOKAI Ghost Local Currency Local Currency U.S.A. Dollars
Japanese Yen
JPY ¥69,702 $434.48
Unit cost (Local Total Cost (Local Total Cost
# Component Part name Autor Source Quantity Currency) Unit cost (US Dollars) Currency) (USA Dollars)
1 Main Microcontroller Teensy 4.0 PJRC
Raspberry Pi digikey.jp 1 ¥4,285 $26.71 ¥4,285 $26.71
2 Image Processing SBC Raspberry Pi 5 (8GB) Foundation akizukidenshi.com 1 ¥17,600 $109.70 ¥17,600 $109.70
3 Sub Microcontroller Seeed XIAO RP2040 Seeed Studio akizukidenshi.com 2 ¥980 $6.11 ¥1,960 $12.22
4 Camera Module IMX219-120 Waveshare waveshare.com 1 ¥2,700 $16.83 ¥2,700 $16.83
5 Line Sensor LBR-127HLD Letex Technology akizukidenshi.com 16 ¥75 $0.47 ¥1,200 $7.52
6 Color Sensor S9706 Hamamatsu Photonics akizukidenshi.com 2 ¥600 $3.74 ¥1,200 $7.48
7 ToF Sensor Module Custom VL53L0X Module Takumi / JLCPCB jlcpcb.com 6 ¥788 $4.91 ¥4,728 $29.46
8 Micro Switch SS-10GL13 OMRON akizukidenshi.com 2 ¥160 $1.00 ¥320 $2.00
9 Tactile Switch DTS-63-N-V Cosland akizukidenshi.com 2 ¥15 $0.09 ¥30 $0.18
10 RGB LED WS2812B Worldsemi akizukidenshi.com 6 ¥50 $0.31 ¥300 $1.86
11 IMU BNO055 Bosch Sensortec
SUNHOKEY akizukidenshi.com 1 ¥2,450 $15.27 ¥2,450 $15.27
12 Display SSD1306 128x64 Electronics akizukidenshi.com 1 ¥580 $3.61 ¥580 $3.61
13 Drive Servo Motor STS3032 Feetech Robotics akizukidenshi.com 4 ¥4,800 $29.92 ¥19,200 $119.68
14 Rescue Servo Motor SG90 TowerPro akizukidenshi.com 2 ¥650 $4.05 ¥1,300 $8.10
15 Rescue Servo Motor SG92R TowerPro akizukidenshi.com 3 ¥780 $4.86 ¥2,340 $14.58
16 Camera Servo Motor GH-S37D Generic GH-S37D 1 ¥580 $3.62 ¥580 $3.62
17 Main PCB Custom made Takumi / JLCPCB jlcpcb.com 1 ¥5 $0.03 ¥5 $0.03
18 Line Sensor PCB Custom made Takumi / JLCPCB jlcpcb.com 1 ¥5 $0.03 ¥5 $0.03
19 Power PCB Custom made Takumi / JLCPCB jlcpcb.com 1 ¥20 $0.12 ¥20 $0.12
20 Rescue PCB Custom made Takumi / JLCPCB jlcpcb.com 1 ¥1 $0.01 ¥1 $0.01
21 Front PCB Custom made Takumi / JLCPCB jlcpcb.com 1 ¥2 $0.01 ¥2 $0.01
22 Side PCB Custom made Takumi / JLCPCB jlcpcb.com 2 ¥1 $0.01 ¥2 $0.02
23 Frame Metal Parts Custom made Takumi / JLCPCB jlcpcb.com 1 ¥30 $0.19 ¥30 $0.19
24 Metal Gear (Motor Side) Custom made Takumi / JLCPCB jlcpcb.com 4 ¥19 $0.12 ¥76 $0.48
25 Metal Gear (Wheel Side) Custom made Takumi / JLCPCB jlcpcb.com 4 ¥32 $0.20 ¥128 $0.80
26 Main Chassis Components Custom made Takumi 3D Printer 1 ¥2,400 $14.96 ¥2,400 $14.96
27 Wheels Custom made Takumi Silicone Rubber Molding 4 ¥90 $0.56 ¥360 $2.24
28 Fasteners Screws, nuts, spacers, etc. wilco.jp 1 ¥2,200 $13.71 ¥2,200 $13.71
29 Wiring Cables, connectors, etc. akizukidenshi.com 1 ¥1,200 $7.48 ¥1,200 $7.48
30 Li-Po Battery 11.1V 2200mAh 30C Fullymax Battery goohobby.com 1 ¥2,500 $15.58 ¥2,500 $15.58
Name of the Total Cost Total Cost ($)
Team name: TOKAI Ghost Local Currency Local Currency U.S.A. Dollars
Japanese Yen
JPY ¥0.00 $0.00
Cost (LocalCost (USA
# Name Software's Tool/Library Description Source Autor Currency) Dollars)
1 Open CV Library Image processing library OpenCV - Biblioteca abierta de Com Opencv.org FREE FREE
2 Numpy Library Numerical computing library NumPy Numpy developers FREE FREE
3 Pyserial Library Serial communication library https://pythonhosted.org/pyserial/ Pyserial developers FREE FREE
4 PyTorch Library Deep learning framework https://pytorch.org/ Meta Platforms FREE FREE
5 Picamera2 Library Camera control library for Raspberry Pi Picamera2 Docs Raspberry Pi Foundation FREE FREE
6 ultralytics Library YOLO-based object detection library https://docs.ultralytics.com/ Ultralytics FREE FREE
7 Albumentations Library Image augmentation library https://albumentations.ai/ Albumentations Team FREE FREE
8 Adafruit SSD1306 Library OLED display driver library https://docs.arduino.cc/libraries/ada Adafruit Industries FREE FREE
9 Adafruit GFX Library Library Graphics rendering library https://docs.arduino.cc/libraries/ada Adafruit Industries FREE FREE
10 Adafruit BNO055 Library IMU sensor interface library https://docs.arduino.cc/libraries/ada Adafruit Industries FREE FREE
11 Adafruit NeoPixel Library RGB LED control library https://docs.arduino.cc/libraries/ada Adafruit Industries FREE FREE
12 SCServo Library Serial servo motor control library https://docs.arduino.cc/libraries/scse Feetech Robotics FREE FREE
13 pololu VL53L0X Library Time-of-flight distance sensor library https://github.com/pololu/vl53l0x-ardu Pololu Corporation FREE FREE
14 Visual Studio Code Software's Tool Software Visual Studio Code: edición de cód Microsoft FREE FREE
15 Autodesk Fusion Software's Tool 3D CAD software Fusión de Autodesk | Software basad Autodesk
Software Freedom FREE/EDUCATION FREE/EDUCATION
16 Git Software's Tool Version control system Git - Descargas Conservancy FREE FREE
17 PlatformIO IDE Software's Tool Data annotation tool PlatformIO PlatformIO Labs FREE FREE
18 Label Studio Software's Tool Image annotation tool https://labelstud.io HumanSignal FREE FREE
19 KiCad Software's Tool PCB design software Kicad KiCad Developers Team FREE FREE
3 pages, rendered as images so they load quickly. The text above is the document's own, extracted from the PDF.
Source code
The team's own source code, 24.7 MB. It is a download rather than part of this page, because a zip is something you open on your computer. It comes from GitHub, which some school networks block.