search

leggedrobotics / icp_localization

This package provides localization in a pre-built map using ICP and odometry (or the IMU measurements).
https://rsl.ethz.ch
BSD 3-Clause "New" or "Revised" License
245 stars 48 forks source link

ICP localization doesn't work as expected #7

Closed JackFrost67 closed 2 years ago

JackFrost67 commented 2 years ago

I'm testing this package to localize my mobile base on a map. I have used AMCL, but I'm struggling with a jumping pose estimation. I had hopes that this package would resolve the problem, but I'm having some problems. My mobile base has bad odometry, so I would expect a correction because of a match between the scan (in point cloud format) and the map (in point cloud format). What I see is no correction, and I can understand why. The map was made using gmapping and later transformed into a point cloud using a simple script I have made.

The result I have: screenshot

jelavice commented 2 years ago

Hi JackFrost67,

are you sure that the icp is running? This is the trajectory from the odometry I assume? I never tested this package in 2D so there might be problems with z or roll and pitch. Did you look at the full pose of the robot? e.g. if the z is too high, the icp would not find any correspondences.

JackFrost67 commented 2 years ago

Hi, thanks for the response.

are you sure that the icp is running?

I mean, looking at the odometry out of the icp, I can surely say it is not running. I managed to align the frame of the scan with the map pcd, but probably is not enough, and so I have no match (resulting in no transformation applied). For the moment, I found another package that works well, I will return on this later on. Thanks for the attention.

arjunskumar commented 2 years ago

@JackFrost67 could you share the details of the package that was working for you?

JackFrost67 commented 2 years ago

@JackFrost67 could you share the details of the package that was working for you?

Hi, I don't want to move people away from this package. I think it works really well for 3D problems, and probably better in 2D, but I didn't have much time to work on it. In any case, I choose iris_lama, which is an ICP + likelihood map (there is a paper on it), and for my problem (which is a simple straight planner) works really great.

arjunskumar commented 2 years ago

Thanks @JackFrost67