A lecture by Gokhan Koca, Koluman Automotive Industry
This project focused on the development of an autonomous sweeping truck designed to carry out predefined cleaning tasks within a controlled and closed operational area. The main goal of the system was to enable the vehicle to perform its sweeping mission autonomously, reliably, and safely, without requiring continuous driver control during operation. Since the vehicle was intended for low-speed cleaning activities rather than high-speed road driving, the system architecture was designed specifically around the operational needs of a sweeping truck.
One of the most important characteristics of this vehicle is that its movement is powered by a hydrostatic drive system. In other words, the vehicle does not rely on a conventional pedal-by-wire architecture for autonomous motion control. Instead, its low-speed movement is achieved through hydrostatic driving mode, which is especially suitable for sweeping operations because it provides smooth, controlled, and precise motion. This type of drive system is highly advantageous in applications where the vehicle must move steadily and accurately at relatively low speeds, typically around 10 to 15 km/h, while continuously performing a cleaning task. For this reason, the hydrostatic structure formed the foundation of the vehicle’s autonomous movement capability.
In order for the truck to move autonomously inside a closed area, it first needed a way to understand where it was located at any given moment. This was achieved through a localization system based on a previously generated point cloud map of the operational area. A point cloud can be described as a three-dimensional digital representation of the environment, created from a very large number of spatial points. Each point corresponds to a real object or surface in the environment, such as the road, curbs, walls, barriers, or surrounding structures. When all of these points are brought together, they form a detailed 3D map of the working area. This map serves as a reference model that the vehicle can use to recognize its location during operation.
To compare its current surroundings with this reference map, the vehicle was equipped with LiDAR sensors mounted on both the front and rear sides. LiDAR sensors measure distances by sending laser beams and detecting the reflected signals from surrounding objects. During operation, these sensors continuously scanned the environment and generated real-time spatial data. The live LiDAR scans were then matched with the previously prepared point cloud map. In simple terms, the system constantly compared what the vehicle was “seeing” at that moment with the map it had already learned before. By finding the best match between the live scan and the stored map, the system could determine the vehicle’s current location within the closed area.
This localization process was further strengthened by integrating additional motion and positioning data. The IMU provided information related to acceleration, rotation, and orientation of the vehicle. The GNSS sensor contributed positioning support, while the vehicle speed data obtained directly from the truck helped estimate how far the vehicle had moved over time. Each of these sensors alone has limitations, but when combined together they create a much more reliable localization solution. Through this multi-sensor structure, the system was able to calculate the vehicle’s position and displacement more robustly within the predefined operational zone.
In addition to knowing its location, the vehicle also needed to understand its surrounding environment in a more meaningful way. For this purpose, a stereo camera was installed at the front of the truck. A stereo camera uses two visual inputs to perceive depth and better interpret the scene ahead. With this sensor, environmental features could be analyzed and segmented, meaning that the system could separate and identify relevant parts of the surroundings more effectively. This helped the vehicle interpret the operational area not only as raw geometric data, but also as structured visual information.
To improve overall perception performance, the outputs of the stereo camera were combined with the LiDAR data using sensor fusion techniques. This approach allowed the system to benefit from the strengths of both sensor types. LiDAR provided accurate spatial and distance information, while the stereo camera contributed richer visual understanding of environmental features. When these two sources were fused together, the vehicle gained a more dependable and complete understanding of the working area. This improved both localization quality and environmental awareness during autonomous operation.
By combining hydrostatic low-speed vehicle control, LiDAR-based map matching, IMU/GNSS/vehicle-speed-supported localization, and camera-LiDAR sensor fusion, the truck was able to autonomously perform its predefined sweeping task within the closed area in a smooth and controlled manner. The vehicle could maintain awareness of its position, interpret the surrounding environment, and continue its cleaning mission with high stability and without operational stress. As a result, the project demonstrated that a hydrostatic sweeping truck can be successfully autonomousized for controlled-area cleaning operations through the integration of mapping, localization, perception, and sensor fusion technologies.
