robot_localization - gps and imu minimal example not working

edit

I have some trouble getting robot_localization working. I have tried several configurations but nothing gives me a good result. So I thought I would try to get the minimal working. I have gps data and imu data. I also have wheel odometry. All this is simulated and all the values are currently almost ideal (no or very small noise). I use navsat with the following conf:

navsat_transform:
  frequency: 30
  delay: 3.0
  magnetic_declination_radians: 0 # 
  yaw_offset: 0
  zero_altitude: true
  broadcast_utm_transform: true
  publish_filtered_gps: true
  use_odometry_yaw: true
  wait_for_datum: false

And then I use one ekf node with the following conf:

  odom0: odom
  odom0_config: [false, false, false,
                 false, false, false,
                 true, true, true,
                 false, false, false,
                 false, false, false]
  odom0_queue_size: 10
  odom0_nodelay: true
  odom0_differential: false
  odom0_relative: false

  odom1: odometry/gps
  odom1_config: [true,  true,  false,
                 false, false, false,
                 true, true, true,
                 false, false, false,
                 false, false, false]
  odom1_queue_size: 10
  odom1_nodelay: true
  odom1_differential: false
  odom1_relative: false

  imu0: imu/data
  imu0_config: [false, false, false,
                true,  true,  true,
                false, false, false,
                true,  true,  false,
                true,  true,  true]
  imu0_nodelay: false
  imu0_differential: false
  imu0_relative: false
  imu0_queue_size: 10
  imu0_remove_gravitational_acceleration: true

The conf is pretty taken from the examples.

I have static transform for map -> odom (they are exactly the same). ekf publishes odom -> base_link transform. The latitude-longitude are exactly aligned with y and x axis.

Wheel odometry is in odom topic.

The problem I am facing: when there is no movement, the filtered odometry (also the transform from odom to base_link) starts moving slowly (the robot itself is not moving). When the robot is moving, everything gets messy pretty quickly.

I have tried to exclude imu, odom, change covariance, all the other parameters. What is my problem here? I would like to get a minimal example working, which would mainly use GPS.

EDIT:

To be exact, the movement of the transform grows strangely. Here are some example of my tf topic (two consecutive messages, we use base_footprint instead of base_link):

header: 
  seq: 0
  stamp: 
    secs: 7
    nsecs: 200000000
  frame_id: odom
child_frame_id: base_footprint
transform: 
  translation: 
    x: 0.00408947608922
    y: -0.0837722308319
    z: 0.0
  rotation: 
    x: 0.0
    y: 0.0
    z: 0.241644993047
    w: 0.970364723872


header: 
  seq: 0
  stamp: 
    secs: 7
    nsecs: 250000000
  frame_id: odom
child_frame_id: base_footprint
transform: 
  translation: 
    x: 0.215371691561
    y: -0.862578741381
    z: 0.0
  rotation: 
    x: 0.0
    y: 0.0
    z: -0.574137811256
    w: 0.818758678541

And after that it goes wild. The strange thing here is that the rotation also changes. Although imu sensor simulator should be pretty stable. And every time I run the simulator, it gets about the same results (about the same time).