Coverage Planners¶
Three path planners have been compared:
Boustrophedon Coverage
Morphology-based Skeleton Coverage
Spanning Tree Coverage
The Boustrophedon coverage planner package had compatibility issues with other packages used for navigation, which meant that Boustrophedon did not work on the MIRTE Master.
The morphology-based skeleton coverage planner performed well in laboratory-like environments but worse in larger open spaces. This planner allowed the robot to maneuver efficiently between objects such as tables and chairs while still being able to get into tight places. The morphology-based skeleton coverage planner also resulted in the fewest sharp turns, which, given the simple odometry of the MIRTE Master (wheel encoders only), helped the robot keep track of its position and thus keep the map clean. However, the skeleton planner performed slightly worse in terms of coverage performance. This was due to the lower density of the planned path, which sometimes caused the robot to miss the edges of larger free spaces that needed to be scanned for objects.
Figure 1:Picture of the skeleton path in a laboratory environment (RViz).
The spanning tree coverage planner performed well in large open spaces, but unlike the morphology-based skeleton coverage planner, it struggled more with planning paths in tighter spaces such as laboratory environments, resulting in separate path loops being stitched together and, possibly, tighter spots between the loops being missed. Also, because of the nature of the coverage pattern, many small-radius turns are generated, which sometimes caused the robot to slowly accumulate drift in the odometry, resulting in less accurate localization and a less accurate map.
Figure 2:Picture of the spanning tree path in a laboratory environment.
Object Localization and Classification¶
For object detection, a localization method and a classification method were developed. Our object localizer relies on bounding boxes that are retrieved from point clouds generated from the depth camera’s depth images. Our object detector was able to find objects sized from 1 cm to 6 cm. A bounding box drawn around a found object in the point cloud is shown in the figure below. The custom-made YOLO26 object-detection model worked as intended, correctly classifying the localized objects. The figure below shows a frame from the RGB sensor of the depth camera in the bottom-right corner.
Figure 3:Bounding boxes drawn around found objects, with YOLO26 running simultaneously.
Manipulation¶
Movement accuracy¶
The manipulation of the actuator has seen major improvements regarding movement accuracy. The initial accuracy of the actuator was around 15 cm. After the code was reworked, the actuator was controlled by actual joint values instead of approximations. This reduced the error significantly, resulting in a movement error of only 1 mm.
Gripping performance¶
The gripping performance of the original robot was slightly improved by adding rubberized patches to the gripping surfaces of the gripper. This has significantly helped the gripper hold the smooth, plastic testing objects when it approached them at a sub-optimal angle.
Figure 4:Figure showing rubberized gripping surfaces.