Prediction step and IMU orientation have little impact on robot_localization position

edit

Hi, I have Odometry (from wheel encoders) and IMU data available, but for each setup the EKF Position replicates the Odometry Position (see fig x_y_offsets.png). At the end of driving several meters the difference is not more than 1cm for x and y value).

The Odometry data only provides absolute X and Y Position and Orientation, which I all fuse. The IMU has in theory absolute orientation, angular_velocity and linear acceleration values. However I want to use the ~use_control parameter so that the prediction step is done based on the velocity commands (doing this for research project). Therefor I cannot fuse any linear acceleration data because that overrides the ~use_control parameter as stated in the documentation. That's fine with me. So the IMU cannot alter the current position, but it provides addition data for orientation of the robot (see attached picture yaw_offsets.png). Looking at the data I see that EKF-yaw value is in the middle between odom-yaw and imu-yaw (both have similar variances).

So I would expect, that over time the EKF and the Odometry position drift apart, since the EKF has a more accurate orientation, which is to my knowledge included at the prediction step.

For variances I made some measurements and calculated the following on my own:

process_x:      3.80192200351542e-05
process_y:      0.000113269151605415
prozess_yaw:    0.000734031281740257
odom_x:         6.59418396241594e-06
odom_y:         0.00011676237024532
odom_yaw:       0.00101425160795909
imu_yaw:        0.00117190176298063
imu_yaw_vel:    0.0154674185084969

I also varied Initial_covariance_matrix values (x,y,yaw) from between 10^6 to 0.001. For low values it took more time steps until EKF was very close to Odometry data again, but no other impact. I also changed some variances to see if they have a biger impact. ie. when increasing the odom_yaw variance I see that EKF-Orientation is always close to IMU-yaw value (which is expected). When increasing the odom_x, and y variance I can see slightly more variation. e.g. when keeping the process variance but putting odom_x, and odom_y variance to 0.5 I can see that the positions vary slightly more (about 5-10 cm) as can be seen in Fig. x_y_offsets_odom_0_5.png. However comparing the last EKF-position with the ground truth (measured by accurate device) I cannot say that the position is more accurate than raw Odometry position.

I would appreciate any help/suggestions. Maybe I have just reached the limit with my sensor configuration, and there is nothing to better. I also considered to include IMU-linear acceleration data and comment parts of robot_localization (so that ~use_control is still active) but I find this rather risky. If you need more information, please tell me.

PS: As I am new to ROS-community, I cannot add pictures yet. As soon as I find out how to generate more Karma I will upload the promised 3 pictures.

frequency: 20

sensor_timeout: 0.2

two_d_mode: true 

transform_time_offset: 0

transform_timeout: 0.1    

publish_tf: true 
publish_acceleration: false

map_frame: map            
odom_frame: odom           
base_link_frame: base_link  
world_frame: odom           

odom0: odom_new

odom0_config: [true, true, false,
               false, false, true,
               false,  false ...