Hi Krzysztof,
I see here a massive angle error. Could it be, that the odom problem can cause that?
I am calculating the z (yaw) angle from angle vector + w like that:
tf::Quaternion q(gOdometer.pose.pose.orientation.x,
gOdometer.pose.pose.orientation.y,
gOdometer.pose.pose.orientation.z,
gOdometer.pose.pose.orientation.w);
tf::Matrix3x3 m(q);
double roll, pitch, yaw;
m.getRPY(roll, pitch, yaw);
if (isnan(yaw)) {
yaw = 0;
}
If the robot sees the wall heating for the first time, it uses the angle 0 (red line on image).
If it sees the wall heating for the 4th time, it uses 90 degrees (pink line on image).
Could this massive error come from this odom problem?
The blue “line” is the robot movement. It scans the room by twisting around the z axis (starting at x/y=0) and if it gets boared (it knows all parts of the room so far) it move a bit until the view is more interesting for it and starts twisting again, and so on. Actually, not the robot is doing that but the PC which remote controls the robot 
EDIT:
I am doing now a yaw angle correction by myself. It turned out, that the yaw angle is around 20 degrees too big after one 360 twist. So, I am caring about my own yaw angle by added raw yaw delta values corrected by a correction factor. This leads to a almost correct angle representation in my machine learning software. The objects are now at the almost correct place in the picture.
It would be very nice of your team, if one could look into that issue too.
And, there is another one: The twist path is represented in the odometry data X and Y positions in the wrong way: My robot twists clock wise but the X/Y coordinates are sent counter clock wise. Look at the end of the path in the pciture, it is turning left, instead of right:
But the forward movements are correct. I checked all my data paths if I have a X/Y problem, but there is no. Also the objects in the room are on their correct places for which I am using also the robots X/Y position and the angle to calculate their places.
So: Why is the twist reverse then?
