Recommended way to detect tabletop objects with MoveIt in Hydro?

[Liang-Ting Jiang]

Hi,

I am wondering what is the recommended way to detect tabletop objects and convert them into graspable with MoveIt in Hydro? I want to detect objects without training and recognizing them. I used tabletop_object_detector in the pr2_manipulation stack before in Fuerte, which seems to be obsoleted in Hydro. I saw in a previous post mentioning Object Recognition Kitchen (ORK), but it seems not better than tabletop_object_detector without training. On the other hand, if I continue to use tabletop_object_detector, then I have to use a lot of old style message types. Any thought would be very much appreciated. Thank you!

-LT

[Marcus Liebhardt]

Liang-Ting,

That tabletop object detector you are using in Fuerte has been re-implemented using Ecto[1] and is now part of ORK [2], [3]. It's basically the same thing, but in a nicer dress.

You should give it a try!

Best,

Marcus

[1] http://plasmodic.github.io/ecto/

[2] http://wg-perception.github.io/object_recognition_core/

[3] http://wg-perception.github.io/tabletop/

--
Marcus Liebhardt 
Control Engineer
Yujin Robot
주소: 대한민국 서울시 금천구 가산동 345-30 남성프라자 #601, 153-023.
Address: Door #601, Namsung-Plaza, 345-30 Gasan-dong, Guemcheon-gu, Seoul, 153-023, Republic of Korea
Website: http://www.yujinrobot.com
Email: marcus.l...@yujinrobot.com
Phone: +82-70-46577073

[Liang-Ting Jiang]

Hi Marcus,

I have tried the tabletop object detector in ORK by learning from the config file you shared in this post: https://groups.google.com/forum/#!searchin/moveit-users/ork/moveit-users/wZe2vq1VzCM/wmCe5PyJ_QgJ

My goal is to get the raw pointcloud of the unknown object clusters so I can do cluster-based grasp planning. However, it seems the pointcloud can only be published by the Publisher after TabletopObjectDetector pipeline. Since these clusters are not really recognized object, i.e., no object_id, poses, I wasn't able to get the pointcloud data published. Is there any way you know to get the clustered pointcloud data without hacking into the ORK code? Attached is the ORK config file I used, and the sample output after the pipeline looks like this:

recognized_objects: 

  header: 

    seq: 10

    stamp: 

      secs: 1380764670

      nsecs: 704890996

    frame_id: /head_mount_kinect_rgb_optical_frame

  objects: []

  cooccurrence: []

Time for 1 detection: 0.264250994

screenshot:

Thanks again for your help!

-LT