MTC near-offline planning and online execution without re-planning.

[jakobbak]

Preface

First a big thank you to all authors of MTC for what seems to be an amazing framework, for doing a lot of things in ROS2 that otherwise seems difficult / daunting.

I will also preface this discussion with acknowledging that my understanding of ROS2 and MoveIt is still in early stages - so please pardon incorrect use of terminology or concepts - pointers and corrections are welcome and appreciated.
I have been spending quite a bit of time over the past couple of weeks doing tutorials, reading the documentation accessible, and looking into the source code, but despite my efforts I have not figured out how to do the below. Hope this is an appropriate place to reach out.

Introduction

I am trying to implement a Pick and Place mechanism for a UR5e, with a vacuum gripper end-effector, using MTC.
My ideal scenario at this point, is being able to define a set of poses for Picking and Placing, calculating the trajectories between these known poses, and then subsequently execute these trajectories on a need-to basis, with an external scheduler telling the MTC node when to execute each trajectory.

The catch is that I want to only calculate the trajectories once, at startup of the MTC Node, and not recalculate them online, as I need to make sure they always gets executed in exactly the same way each time they are repeated.

Setup

I am on Ubuntu 24.02 with mtc branch jazzy cloned into ~/ros2_ws/src/.
I have a sufficiently detailed URDF and SRDF setup and accompanying config and launch files in suitable packages robot_description and robot_movit_config.
Pick and Place poses are initially defined in SRDF, but will at a later stage be passed as parameters, and eventually as messages coming from another Node - but these latter aspects are not part of this discussion.

What is working so far

Creating a MTCNode based on the example in the official MTC Pick and Place tutorial, but without grasps, and using my own set of poses, and of course a different robot.
I've created:

• one task,
• with a Parallel container (moveit::task_constructor::Alternatives)
• which contains four Serial Containers (moveit::task_constructor::SerialContainer),
• each containing a number of stages, that corresponds to moving from a Home position to one of two positions and back to Home, through a number of intermediate positions, all using the Cartesian solver
• there is also a stage in the beginning of the task that takes the robot from its Current Position to Home using OMPL.

I can plan this, and I can have it execute with auto result = task_.execute(*task_.solutions().front()), in which case it executes the first move from Current Position to Home, and then the first stage in the parallel container, the former itself being a serial container of stages.

(In the GIF we see RViz receive solutions from MTC and execute the first two stages. I then manually collapse the four Serial Containers and select them one by one, select their solution, and use the Execute Solution button in RViz to execute them.)

Planned work

From here I plan to first attempt to execute in turn (by separate calls to task_.execute) the four parallel stages, as they each have the same start and end position. Afterwards I hope to implement a service in the node that can parse a number or string and select the corresponding stage in my parallel container to execute it.
To do this I need to be able to select the appropriate solution in task_.solutions() as it is used in the call auto result = task_.execute(*task_.solutions().front()).

Questions

Selecting a specific solution

From most of the tutorial material I have encountered I only see *task_.solutions().front() being used when executing.
Is there a way to pick one of the other solutions by name or number?
I realise my limited experience with STL shines through here, but I tried using *task_.solutions().find("Pick Record from Conveyor") for example, but encounter compilation errors.
Is the intention (or consequence, which is fine) that the user of MTC write their own method for finding the appropriate solution? (Perhaps something along the lines of what is described here - note though the requirement for overloading operator==).

Just use multiple tasks instead

Should I give up on using one task with parallel solutions, and instead use multiple tasks, seeing that they are all independent from each other any way? Each trajectory will have requirements for start positions, but that would be a controller problem anyway, not an issue for telling MTC to execute a particular task, which are easily named, and hence able to get connected with a service or an action.
I will test this out next, while I await any comments on the above.

Am I using MTC completely wrong?

Is it the wrong use pattern for MTC?
When I generate a set of parallel solutions, then have an external scheduler (a state machine I'll create for the purpose) executing them through MTC, but without recalculating anything, is that a "waist" of MTC? Or goes against its intent?
I gravitated towards MTC because it seemed that it was an "all-in-one" package, had good tutorials, and was actively supported. I found it hard to get an overview of other packages / frameworks to use within MoveIt, and got the impression I would have to write a lot of nodes from scratch to get a similar functionality.
But since the solutions are published, perhaps there is an easier way to select them and execute them programmatically from somewhere they also exist? (the planning group?)

[rhaschke]

Indeed, you should instantiate 4 separate MTC tasks and one additional to move from start to home.
A solution's task is intended to be executed entirely, not partially.

Is there a way to pick one of the other solutions by name or number?
I realise my limited experience with STL shines through here, but I tried using *task_.solutions().find("Pick Record from Conveyor") for example, but encounter compilation errors.

task_.solutions() returns a vector-like object, not a map. Hence you can refer to solutions via task_.solutions().front(), task_.solutions()[0], task_.solutions()[1], etc. However, solutions are always sorted by cost and you want to pick the most cost-effective solution, i.e. the first one.

Is that a "waste" of MTC? Or goes against its intent?

No. Your approach is perfectly valid. Just use separate tasks.

[jakobbak]

Thank you so much for your quick reply. I will go ahead with the multiple-task approach then.