Add Prediction Interface and linear position prediction model to Object Memory

requested review from @RainerKartmann

assigned to @phesch

Thanks. Could you fix the problem with the inverse result?

added 1 commit

c78693b4 - Minor changes to prediction code

Looks good! I made some changes, but more on the style level.

As a refactoring, we could move the static function Eigen::Vector3d makeLinearPrediction() out to some general place. Maybe even SimoxUtility?

added 3 commits

ee30cbb7 - Move linear prediction code to simox
6ff44348 - Shift timestamps around current time to avoid floating point imprecision
ba6e6a5c - Merge branch 'linear-prediction-from-simox' into 'linear-prediction'

Compare with previous version

added 1 commit

685ddf54 - Update to change in Simox

Compare with previous version

I generalized the regression code in Simox. This changed the header and name of the regression class, so I updated the branch. I hope this won't create too many conflicts for you, as it is just a small change :)

added 5 commits

5f895988 - Make object pose prediction lookback a parameter
4ee022c4 - ObjMemory,Segment: Remove uses of armem::fromIce
e75ac32e - Make getAvailableEngines return map instead of vec
c102907d - MemoryViewer.cpp: Replace uses of armem::fromIce
d16da192 - Add ObjectPose pred. to MemoryInterface

Compare with previous version

added 4 commits

5104c3b5 - Implement suggestions from Gitlab MR
bc2aad20 - Add time offset and window params to ObjMem Visu
864034c7 - Add fallbacks to linear prediction for edge cases
ea64a429 - Make ObjPosClientEx visualize multiple predictions

Compare with previous version

resolved all threads

Some things I noticed while working on this MR:

Using std::stringstreams for error messages in BOs is not as great as we thought because it makes the BOs non-copyable, so you can't use them in vectors anymore.
Last time, we discussed that predictions should use MemoryIDs that are resolved to the snapshot level. However, I think that object pose predictions require instance indices for disambiguation. In particular, the current code only works if every entity in the instance segment only has a single instance per snapshot. If it has multiple object instances, the predictions will mix up the different objects' poses. I'm not quite sure how to resolve this since ObjectIDs support arbitrary strings as instance names - armem only supports integers, so the conversion would be a little iffy.

Follow up:

Refactor prediction code from source/RobotAPI/libraries/armem_objects/server/instance/Segment.cpp into the library RobotAPIArmarXObjects (i.e. something like: ObjectPose predict(vector<ObjectPose> poses, time))