Perception is responsible for object detection, classification, and spatial localization. In this study, two object-detection and localization approaches were evaluated: a 2D detection and depth-projection approach and a 3D segmentation approach based on point-cloud clustering. For classification, a classical computer-vision approach and a model-based approach were both considered.
Initially, the goal was to distinguish between graspable and non-graspable objects and between electronics and non-electronics. However, because the standard MIRTE Master platform had limited camera quality and limited computational resources, the final scope was reduced to two practical classes: graspable versus non-graspable and colourful versus greyscale. This choice was made to keep the system feasible within the available hardware and to focus the evaluation on the most relevant objects for the laboratory-cleaning task.
A custom YOLO-based detector was trained on a small labelled dataset collected from the onboard camera, while the 3D pipeline used point-cloud filtering, plane removal, and clustering to estimate the position and orientation of candidate objects. The combination of these two perception streams was then used to support the grasping and sorting process.