Bonus: A pick-and-place system

A system of nodes

Real robots split work into nodes that each do one job and agree only on topics:

Any node can be replaced (a camera for /perception, a smarter planner) without touching the others. The planner ignores /targets stamped before the last /pick_done: they were seen before that pick finished, and may still list the ball it took.

Waiting without blocking

A pick takes seconds, but a callback mustn't block. Make it a coroutine: async def on_target(self, msg), with await on every future, such as r = await client.call_async(req). While it waits, the executor runs other callbacks.

There's one catch, the same on a real robot. A coroutine callback keeps its callback group busy until it finishes, and a node's callbacks all share one mutually exclusive group by default. The reply it awaits arrives through its client, which is in that same busy group, so it never gets through: a deadlock. Here a warning names it. Put the clients in another group, such as a ReentrantCallbackGroup(), whose callbacks may run at any time.

Configuration

The launch file sets parameters, which each node declares:

/**:
  ros__parameters:
    cup: [-0.06, 0.34, 0.0]
perception:
  ros__parameters:
    rate: 2.0
picker:
  ros__parameters:
    hover: 0.1
    lift: 0.2
    move_time: 1.0

/** applies to every node. The gripper keeps its allow_stalling from step 7.

Your task

Clear the table: every ball in the cup within 36 s. Publish the targets, choose the ball nearest the cup, give the picker's clients their own group, and write the pick sequence.