robot_localization output is non deterministic?

asked 2019-10-04 02:49:42 -0500

311 ●11 ●16 ●24 http://weeklyrobotics.com/

I recently thought I would be really clever and save time tuning robot_localization live on my robot. What I did was to create something akin to UMBMark where I was driving the robot in a rectangular pattern and stopping the robot at roughly the same position and orientation as a my starting point (0,0).

I made around 5-66 laps through this rectangular path (about 2m x 1m) and I recorded my sources of odometry, IMU and static transforms.

Then I was hoping that by playing back the bag file and running the ekf I was hoping to play with the config and tune my localization.

What I noticed instead is when I run the exactly same setup (using a single launch file) I get different results every time. In 5 of my runs here are the results of a final robot_position when the bag finishes:

(0.338, 0.361)
(0.336, 0.359)
(0.482, 0.320)
(0.394, 0.342)
(0.319, 0.375)

I was expecting that the results will be very close to one another (like the first and the second result) and I was quite surprised to see such a spread of values.

Is there anything that I'm missing?

Comments

Then I was hoping that by playing back the bag file and running the ekf I was hoping to play with the config and tune my localization.

how are you playing back the .bag exactly? Which options to rosbag play are you specifying?

gvdhoorn ( 2019-10-04 05:04:19 -0500 )edit

@gbdhoorn: Here is how I launch it from the launch file:

I also set the use_sim_time param to true like this:

msadowski ( 2019-10-04 05:43:51 -0500 )edit

add a comment

answered 2019-10-04 03:59:59 -0500

Delb

3907 ●16 ●105 ●68

updated 2019-10-04 04:01:16 -0500

That is indeed a big offset in your result but that is not due to robot_localization but rosbag. You can find in the rosbag tutorial a section about the limitation of rosbag :

In the previous section you may have noted that the turtle's path may not have exactly mapped to the original keyboard input - the rough shape should have been the same, but the turtle may not have exactly tracked the same path. The reason for this is that the path tracked by turtlesim is very sensitive to small changes in timing in the system, and rosbag is limited in its ability to exactly duplicate the behavior of a running system in terms of when messages are recorded and processed by rosrecord, and when messages are produced and processed when using rosplay. For nodes like turtlesim, where minor timing changes in when command messages are processed can subtly alter behavior, the user should not expect perfectly mimicked behavior.

So tiny changes in the timestamp of each messages will affect the final localization since robot_localization fuse data from different sensors based on their timestamp. If timestamps differ between two tests then the ouput will differ two.

NB : I'm not saying that robot_localization output is deterministic (as I can't tell), just that the input data wasn't exactly the same each time so the ouput data can't be the same too.

edit flag offensive delete link

Comments

I'm not entirely sure this is correct. geometry_msgs/Twist (ie: input to turtlesim) is not a stamped message (so doesn't contain a Header field). This means that the only time information available to turtlesim is the time at which it receives a message. As rosbag play does not reproduce the exact (inter-message) timing from when it was recorded, this causes what is described on the wiki page you quote.

robot_localization consumes mostly *Stamped variants of messages, or message types that include header directly (ie: nav_msgs/Odometry, sensor_msgs/Imu, geometry_msgs/PoseWithCovarianceStamped, and geometry_msgs/TwistWithCovarianceStamped).

rosbag play should not replace the stamps in the messages from the .bag, and if reception time (or inter-message delta-T) is not used by robot_localization, the section you've quoted may not explain what the OP is seeing.

gvdhoorn ( 2019-10-04 05:02:59 -0500 )edit

You are right about the stamps not being modified I've made a hasty conclusion after seeing this in the robot_localization wiki :

Continuous estimation. Each state estimation node in robot_localization begins estimating the vehicle’s state as soon as it receives a single measurement

That's what made me say that reception time is used by the robot_localization, am I wrong ?

Delb ( 2019-10-04 05:16:06 -0500 )edit

I believe that just means that as soon as robot_localization receives a message on any of the configured topics it will start updating and evaluating its internal model.

But I only commented on the difference between turtlesim and robot_localization. I haven't checked any code, so it could well be that 'normal' time, and differences in it, still affects the output.

gvdhoorn ( 2019-10-04 10:00:35 -0500 )edit

add a comment

robot_localization output is non deterministic?

Comments

1 Answer

Comments

Question Tools

Stats

Related questions

robot_localization output is non deterministic? edit

Comments

1 Answer

Comments

Question Tools

Stats

Related questions

robot_localization output is non deterministic?