Currently I am using rtabmap using wheel odometry and 3D lidar.
I am trying to extend the map I have. During that session I face two issues.
1) When there is an angular command the localisation drifts so the mapping gets screwed up. Not sure which parameter to tune to avoid drifting of localisation.
I will attach the link for the video regarding this issue.
2) I did a small hack for now for first issue to stop mapping when I turn. Just map in the straight line.
while trying to do mulit session mapping. The occupancy grid which was mapped previously started to disappear, slowly. Not sure why that occurs also dont really know which parameter I need to take a look.
I will attach the link for the video, config and 2d map as well.
1) Is this in simulation or in real environment? With a real 3D lidar sensor (I assume it is ring-like lidar), the lidar should be unskewed before sending the point cloud to rtabmap. When the robot turns, the lidar point cloud is likely to be skewed, thus this can cause not so accurate localization. If you have a velodyne, you can set a fixed frame (e.g., your wheel odometry) in velodyne launch file to do the deskewing. Otherwise after looking the videos, it seems that your wheel odometry is drifting more for angular motion.
2) If you don't deskew the lidar, it is a good idea to ignore those scans when the robot is turning. With some parts of the map disappearing, it seems like "Rtabmap/TimeThr" is set. However, looking at the video and your database, Rtabmap/TimeThr is already 0 and I don't see parts of the map disappearing. If you refer to obstacles disappearing, this is when new local occupancy grid are created in same area, they will override the obstacles and empty space. For example, if an object has been moved between the two traversals, it can be cleared when to robot "sees" that the obstacle has been removed.