ompl_interface::ConstrainedGoalSampler::sampleUsingConstraintSampler doesn't appear to sample?

[werner291]

Hello,

I was working on trying to implement a custom constraint_samplers::ConstraintSampler, but I noticed that only the project method gets called, never the sample method.

The state I get as input always appears to be the same (I think it's the same as the starting state?) when I print it out.

When reading through https://github.com/ros-planning/moveit/blob/master/moveit_planners/ompl/ompl_interface/src/detail/constrained_goal_sampler.cpp#L90, I also don't really see any source of randomness in there... it appears to just be projecting the same state over and over.

Is this the intended behavior?

[werner291]

Looks like sample can be swapped into the place where project is used with no further changes to the code; it even takes the same arguments. Maybe that was what was supposed to happen there?

[werner291]

Made the swap in my own fork of the code... so far seems to work fine?

[v4hn]

Maybe that was what was supposed to happen there?

project was apparently just treated as another interface for sampling here where the first parameter serves as input and output.
From what I can see, the MoveIt code base only uses project() at the moment and if you change it to sample, then project will be unused. So they are basically redundant interfaces in case some plugin and user code wants to differentiate. Clearly no nice design.

I'm open to fix the semantics here. If you want to change it to

constraint_sampler_->sample(work_state_, work_state_, planning_context_->getMaximumStateSamplingAttempts()))

please file a pull-request.

[werner291]

Well, the semantics seem quite different, don't they?

The way the documentation is written, "sample" generates a completely new robot state, writing it into the passed-in data-structure, whereas "project" actually takes the passed-in structure, and applies an operation to it that makes it compliant with the constraints.

Either way, see the new pull request: #2872

[simonschmeisser]

Please also have a look at #2273 It was finished and merged in MoveIt2 but unfortunately not in MoveIt (1) Maybe it could be backported? (If that is what you need in the end ...)

[werner291]

Oooh, that looks fancy! I'll check it out; I've never really been a huge fan of the "carthesian planning + IK" approach, since IK is kinda meaningless for my particular robot (base is very mobile) without directly tying it to the planning process.

At the moment though, I'm just using a constrained sampler, which seems to work just fine for my purpose.

I still believe my PR (#2872) would be useful, since in my particular sampler, projecting and sampling really are two distinct actions; one relies on the other, actually. (It produces samples uniformly, then projects them to obtain constrained ones.)

I can put all of my code in the project, method, but that feels kinda wrong, and I'd prefer not to depend on what appears to me to be a bug in MoveIt.