Goal: Design and fabricate a low-cost roller-coaster track inspection robot capable of collecting high-fidelity images and other relevant data to safely evaluate structural conditions.
Students:
Zachary Scullin: zachary.scullin@yale.edu
Thomas Venarde: thomas.venarde@yale.edu
George Ploumis: george.ploumis@yale.edu
Brandon Lin: brandon.lin@yale.edu
This capstone project aims to create an autonomous robotic inspection system capable of traversing roller coaster tracks to collect visual and structural data to inform track inspection and maintenance efforts. Our robot uses a custom-designed chassis and bogie system to traverse straight, curved, and inclined track sections. It then creates a 3D scan of the inspected track from which bulk defects and small-scale defects can be observed.
The safety and functionality of modern-day roller coasters depend greatly on frequent inspections of the track structure, welds, and general mechanical components. From our literature review and discussions with Engineers from Vekoma, a roller coaster manufacturer, and Lake Compounce amusement park, we found that current inspection processes are mostly manual. Maintenance workers climb along the track structures to visually examine all parts of the coaster (welds, bolts, structural pieces of the track), adding a safety risk. The primary alternative to manual inspection is drone inspection, which has its own complications like cost and weather dependency. Inspections range from daily to monthly to yearly and contribute to track downtime.
The final prototype featured a custom suspension system inspired by real roller coaster wheel assemblies, a motorized drivetrain, and onboard data collection. To test the robot, we designed and 3D printed our own roller coaster tracks, including both straight and curved sections, since access to real coaster tracks was unavailable. The completed curved track consisted of 22 individual sections and required more than 12 days of continuous printing.
To inspect tracks, we implemented advanced 3D mapping and SLAM (Simultaneous Localization and Mapping) technology using onboard cameras, sensors, and a Raspberry Pi computer. This allowed the robot to generate detailed 3D reconstructions of track sections and track its position autonomously in real time. We also designed the robot to be modular, making it easier to manufacture, assemble, and improve individual components without redesigning the entire system.
Throughout the project, we manufactured custom aluminum and steel components using waterjetting, milling, welding, and 3D printing. We also validated the robot’s mapping capabilities using several SLAM and 3D reconstruction software packages to demonstrate that the system could accurately model track geometry and movement. Overall, the project successfully demonstrated the feasibility of an autonomous roller coaster inspection robot and created a strong foundation for future development and real-world testing.
Shown above are prototypes 1, 2, and 3. Prototype design was an iterative process, with changes being made in the number of wheelsets, wheelset and suspension design, and sensor location over the course of the project. This process was enabled by quick iteration through 3D printing.
The manufacturing process combined waterjetting, milling, welding, machining, and 3D printing to produce the robot’s custom mechanical components. The chassis and structural brackets were machined from aluminum, while the bogie frames were waterjetted from steel and welded together using custom 3D-printed alignment jigs. Precision axles, motor adapters, and drivetrain components were manufactured on a lathe to ensure proper wheel alignment and smooth motion along the track. Additional custom parts, including wheels, belt-drive gears, and track sections, were 3D printed to allow rapid prototyping, modular assembly, and efficient testing throughout the project.
Roller coaster trains are secured to their tracks via undercarriage assemblies called bogies. These typically consist of three sets of wheels known as running, guide, and upstop wheels that support the cart, keep it aligned with turns, and secure it when inverted or climbing inclines. Our device uses two bogie assemblies with a plate chassis. MIG welding was used to fabricate the core plate structure, and a belt drive was chosen for simplicity of assembly. Following rolling resistance verification, a 1:1.5 gear ratio was chosen to gear up the motors for a theoretical lap time of 30 seconds at 60% of the motor’s peak rpm.
CAD Model of Bogie (Left) and Manufactured Bogie (Right)
The electronics system was designed to allow the robot to operate semi-autonomously while collecting accurate mapping and positioning data for SLAM and 3D reconstruction. The robot used an Intel RealSense depth camera to capture images and depth information of the track, while an onboard IMU sensor on an OpenCR board measured acceleration, orientation, and movement data. A Raspberry Pi 5 served as the main onboard computer, handling communication between sensors, processing data, and enabling wireless operation of the robot.
Motor control and sensor synchronization were managed using an OpenCR microcontroller connected to Dynamixel motors, allowing the robot to precisely control movement while timestamping and organizing incoming data streams. ROS2 was used as the communication framework because it could efficiently synchronize information from the camera, IMU, and motor encoders in real time. The electronics architecture was designed to be modular and plug-and-play, minimizing the need for custom soldered circuitry and simplifying assembly and troubleshooting.
To validate the system, several SLAM and 3D mapping software packages were tested, including ORB-SLAM3 and RTAB-Map. These systems successfully demonstrated the robot’s ability to track its movement, reconstruct track geometry, and generate dense 3D point clouds from collected sensor data. The final electronics setup provided the foundation for autonomous inspection and future expansion of the robot’s mapping capabilities.
Electronics Schematic
The final prototype, seen here in CAD form, used two Dynamixel motors and a RealSense Depth Camera, paired with an OpenCR and a Raspberry Pi for control and data handling.
The full SLAM pipeline was successfully tested using a 45-second run of the robot on the scaled roller coaster track. Using synchronized RGB-D camera data and IMU measurements, the system generated both a complete trajectory of the robot’s motion and a dense 3D point-cloud reconstruction of the track. Throughout the test, the visual odometry maintained stable feature tracking and successfully performed loop closure, allowing the software to minimize drift and accurately align the beginning and end of the track loop.
The resulting 3D reconstruction captured the geometry of the track in significant detail, including rail sections and support structures, demonstrating that the robot could create a consistent and accurate map of its environment. To validate the reconstruction, measurements from the generated map were compared to the original CAD model of the track. The reconstructed track length differed from the CAD model by only 2 cm, while the width differed by only 1 cm, corresponding to less than 1% error in both dimensions. These small deviations were considered acceptable given the modular assembly of the physical track and confirmed the overall accuracy and reliability of the SLAM and mapping system.
Testing was done on a 3D printed track, measuring over 8’x5’ and requiring more than twenty separate prints.
The final prototype represented a significant upgrade in durability, transitioning to aluminum hardware from previous 3D printed designs. The wheelsets, or “bogies,” were made from waterjetted and welded sheet metal, while shafts and brackets were machined on the lathe and mill, respectively.
RTAB-Map visual-odometry feature tracking between two RGB frames during the capture. Yellow points are detected feature points, cyan lines connect matched features across frames. Dense, well-distributed correspondences across both the foreground track and the background scene indicate robust IMU-aided odometry throughout the run.
In order to check for track defects, a 3D scan of the track was made using the RealSense depth camera and SLAM: Simultaneous Localization and Mapping. This produced the above 3D scan of the track, while also combining image data and location data from an accelerometer and a gyroscope in the OpenCR to produce a reproduction of the robot’s path. Deviations from the expected path can be used to discover bulk defects, while the 3D scan and images themselves can be used to find defects like corroded bolts or faulty welds.
The team at the final presentations.
Left to right: George Ploumis, Thomas Venarde, Brandon Lin, Zach Scullin
Thank you to:
Final Poster