I need to detect known objects in a workspace and generate very good 6-DOF pose estimates for an automated assembly project. I would prefer a non-neural-network approach, something like 3-D features plus ICP for refinement.
So far it seems the go-to solution in ROS1 was ORK (Object Recognition Kitchen), but this was developed for Indigo and mostly ported to Jade, with a few packages making it to Kinetic. ORK was an extensible framework with several pipelines included so you could quickly try things out.
I am not finding anything comparable in ROS2. The perception metapackage has lower-level tools (image handling, wrappers around OpenCV and PCL, etc.) that would be useful for implementing a pipeline.
Why did such an important functionality disappear from ROS? Could it be that people find it trivial to implement it from libraries like PCL? Maybe research interest has shifted toward machine learning approaches? If I missed something obvious, please let me know.