ROS Resources: Documentation | Support | Discussion Forum | Service Status | Q&A answers.ros.org
Ask Your Question
0

Smearing / Ghosting of laser scan in move_base costmap

asked 2016-04-10 12:31:39 -0500

jordan gravatar image

updated 2016-04-12 15:27:52 -0500

Hello ROS Community,

I'm seeing a problem where laser scans are exhibiting rotational "smearing" when marking the move_base costmaps. This results in "ghost" obstacles in both the global and local maps, and these have the potential to cause path planning to fail. Typically, these smearing effects are cleared out when subsequent scans detect free space, and so the issue does not always cause navigation to fail. However, it does sometimes cause sub-optimal behavior, such as a need to re-plan.

The robot spec:

  • Kobuki base
  • Raspberry Pi 2, running ROS Indigo / Ubuntu ARM
  • Rhoeby Dynamics R2D LiDAR (@ 3 Hz scan rate)
  • Navigation stack (gmapping, move_base)

In the below screenshot: observe the accumulated LaserScans (in white), the local costmap marked areas (in yellow), and the inflated obstacles (in blue). Notice the ghosting of the walls in the top-left part of the costmap (in yellow).

image description

If I set "Decay Time" on the LaserScan topic in RViz to accumulate the scans, and then observe the LaserScan plot over the same time period as when the costmap smearing occurs: I see that there is no such smearing in the decayed LaserScan plot. So the errors in rotation only show up in the costmap, and not on the LaserScans themselves. How/why would that be happening? How is it possible for this difference to exist?

Another observation about this phenomenon is that the problem only occurs when the rotational velocity of the robot is a non-constant, eg. when beginning turning from stopped, or when slowing down from a turn which is in progress. So the problem is a transient one that occurs during theta acceleration. Also of note is the fact that there are no similar smearing effects in translation (robot accelerating along x/y).

The problem is:

  • unique to theta
  • only occurs during theta acceleration (or deceleration)
  • does not seem to affect LaserScans themselves (as plotted in RViz)
  • only appears on the move_base costmaps (global and local costmap)

The fact that theta acceleration must be non-zero, leads me to think the problem could be due to delays in the odom TF, perhaps with a root cause in the Kobuki. Is the Kobuki just slow to reflect changes in it's rotation? Perhaps the odom transform lags the actual robot rotation?

But why would this not also affect the "decayed" LaserScan data as plotted in RViz?

The fact that the LaserScan plot is not affected, leads me to think the error could be rooted in move_base, although I doubt it. Perhaps there is some delay in the uptake of changes in the TF? But why only on rotation?

Thanks for any comment.

ADDENDUM:

I've checked the TF frames, and it seems to look reasonable:

image description

The frames utilized are:

The topic /laser_data is published in tf frame: base_laser (rigidly attached to base_footprint)

The topic /move_base/local_costmap/costmap is published in tf frame: odom

The Global frame is: map

local_costmap:

global_frame: odom robot_base_frame: /base_footprint

global_costmap:

global_frame: /map
robot_base_frame: /base_footprint

ADDENDUM 2:

I made the following change ... (more)

edit retag flag offensive close merge delete

Comments

Have you tried mapping a priori using gmapping and then pairing move_base with amcl on the (now) known map? I'm asking because your costmap configuration might be assuming a static map, not one generated online using SLAM. Conversely, check your costmap configuration.

spmaniato gravatar imagespmaniato ( 2016-04-10 12:43:14 -0500 )edit

Yes, I've done that, ie. taking a previously built map (and then using amcl), but it shows exactly the same behavior. Note: mapping and navigation are essentially work fine, modulo this transient rotational issue.

jordan gravatar imagejordan ( 2016-04-10 13:06:24 -0500 )edit

I see, OK. Then yeah, the next logical suspects are the Odometry messages and the odom TF. Have you checked/ran tf view_frames, tf tf_monitor, roswtf, rostopic echo odom, etc.? Anything suspicious?

spmaniato gravatar imagespmaniato ( 2016-04-10 14:00:49 -0500 )edit

I've checked the TF frames, and it seems to look reasonable (see ADDENDUM:). Appreciate your input.

jordan gravatar imagejordan ( 2016-04-10 16:05:06 -0500 )edit

The plot thickens. What are the values of your global_costmap/robot_base_frame, global_costmap/global_frame, local_costmap/robot_base_frame, and local_costmap/global_frame parameters? (You've already mentioned a few of these in your addendum; just getting the full picture.)

spmaniato gravatar imagespmaniato ( 2016-04-10 20:59:46 -0500 )edit

local_costmap: global_frame: odom robot_base_frame: /base_footprint

global_costmap: global_frame: /map robot_base_frame: /base_footprint

jordan gravatar imagejordan ( 2016-04-11 10:04:51 -0500 )edit

1 Answer

Sort by ยป oldest newest most voted
1

answered 2016-04-12 01:11:17 -0500

ahendrix gravatar image

It looks like the costmap is smearing the laser scan because it doesn't quite correctly account for the amount of time that it takes to actually rotate the scanner and capture the scan. This probably isn't noticeable on faster scanners or slower robots, because the distortion is lower.

rviz is correct, and waits for transforms for the entire duration of the scan to arrive before attempting to transform it.

costmap2d waits for a fixed window of transforms to arrive before attempting to transform the scan. If your scan took longer to acquire than the fixed time that the costmap has hardcoded, then it will attempt to transform before all of the transforms have arrived, and will probably smear the scan.

If you're comfortable building navigation from source, you can manually increase the tolerance here: https://github.com/ros-planning/navig... to something at or above your scan time, rebuild navigation, and see if that helps. (For a 3Hz sensor, a tolerance of 0.34 should be about right).

If that fixes the issue, open a ticket against navigation's bug tracker.

For reference, the section of rviz that computes the correct tolerance is https://github.com/ros-visualization/...

edit flag offensive delete link more

Comments

Thanks very much for the suggestion. I changed the code, but the problem remains the same. Yes, having a faster scan definitely reduces the degree to which this is a problem. And slowing the robot down also mitigates it. It's so curious that the RViz accumulated plot seems fine.

jordan gravatar imagejordan ( 2016-04-12 10:19:55 -0500 )edit

Are you certain that you've rebuilt and are actually running your modified version of navigation, instead of the system-installed version? Did you remember to build your workspace and source the workspace once it was built?

ahendrix gravatar imageahendrix ( 2016-04-12 13:26:33 -0500 )edit

I like to add some kind of obvious print/ROS_WARN when I modify a package that's also installed on the system, and then look for that at run-time to make sure that it's actually running my modified version.

ahendrix gravatar imageahendrix ( 2016-04-12 13:38:24 -0500 )edit

Yes, indeed, I usually put a print to confirm execution, just as you say. Please see ADDENDUM 2. If the problem is TF delay, how would that not affect translation too? Why is only theta acceleration implicated in this, and not constant angular velocity too?

jordan gravatar imagejordan ( 2016-04-12 15:25:08 -0500 )edit

Your Answer

Please start posting anonymously - your entry will be published after you log in or create a new account.

Add Answer

Question Tools

1 follower

Stats

Asked: 2016-04-10 12:31:39 -0500

Seen: 525 times

Last updated: Apr 12 '16