This is one of our most complex projects: an autonomous tracked platform drives alongside a train, and a robot mounted on it points a 3D scanner and cameras at the wheels, axle boxes, springs and side bearers. Each finding is linked to a 3D scan and a photo, and the final decision is made by the inspector. It is a pilot project for rail freight wagon inspection.
The project brings together several fields that are often handled separately: centimetre-accurate satellite navigation, mobile platform control, an industrial robot, 3D machine vision and artificial intelligence.
1. Navigation: our own RTK base station
A standard GPS receiver is off by several metres - not good enough for wagon inspection. So we set up our own RTK base station (based on a Septentrio mosaic-X5) 160-200 m from the platform. It sends corrections to the platform's receiver, and if the corrections drop out for more than a minute, the system automatically switches to a backup VRS network. We measured and corrected the offset between the two sources - they agree to within 4 mm.
- The receiver on the platform is a Septentrio mosaic-H with two antennas 1.28 m apart: it provides not only position but also heading.
- Static for 60 s: 600 out of 600 measurements in RTK Fixed mode, horizontal error approx. 9 mm (1σ), 2 cm (95%). Heading error approx. 0.23° (1σ).
- Satellite reception deteriorates between tall wagons. With a suitable elevation mask (12°), the RTK Fixed share next to the wagons rose from about 50% to 93-98%.
- When the antenna heading drops out briefly, an inertial sensor holds the heading.
More in our article RTK GNSS: centimetre accuracy for mobile robots.
2. The platform drives itself
The route is stored as a smooth curve. Ten times per second, the PC calculates the heading correction (Stanley controller) and sends commands over Modbus TCP to a Delta PLC, which drives the track motors. When following the route, the median deviation is 5.2 cm, and 95% of the time it stays within 12.1 cm.
For stopping at an axle box, an AI model (SAM 3) is being prepared that finds axle boxes in the camera image. On a recorded run, it found 8 out of 8 axle boxes with no false detections.
3. The robot and the 3D scanner
The platform carries a Yaskawa Motoman MH5F robot with Photoneo PhoXi and MotionCam-3D scanners and a Basler HDR camera. We control the robot from the PC via HSES: the PC streams points while the robot is already moving, so the motion is smooth, without stops.
The key point is that the robot path is adjusted automatically from the 3D scan. The platform never stops in exactly the same place, so the scanner first finds the bolt circle of the axle box (fit error 0.43 mm), and the robot poses for the spring are calculated from it. More in robot path correction from a 3D scan.
To make the camera and robot "speak" the same coordinate system, we carried out hand-eye calibration: the RMS error fell from 8.7 to 3.2 mm, and repeatability of the same pose is 1.45 mm (based on few measurements so far).
4. What the 3D scans show
- Three views - one model. We merge scans from three directions into a single model; the merging error (ICP RMS) is 0.12-0.20 mm.
- Wheel diameter. We compared 8 wheels with manual measurement: after correcting a systematic offset, all wheels agreed within ±0.8 mm, and repeat scans of the same wheel differed by 0.2-1.5 mm.
- Shiny metal. Wheels and axle boxes reflect light. We do not use matting spray: we detect and mask glare zones, so all 31 false "defects" caused by reflections were rejected. More in 3D scanning of shiny surfaces.
5. What comes next
We are currently fine-tuning the calibration, the robot motions and the wheel analysis. The next step is testing the stop at the axle box under real conditions. The figures in this article are test results, not a final guarantee for operation.
What this means for you
The same solutions apply elsewhere: outdoor inspection, large objects that have to be driven around, a robot that has to find its exact working position on its own. If your task needs motion, precise positioning and 3D vision together, write to us. More about our machine vision and robot integration services.