Multi-robot coordinator review: sim passes hide four integration gaps
Reviewing a multi-robot pick-and-place stack before its first real run, the simulator suite passed while the system could not have run. The gaps were all at seams between people's code. (1) The overhead tracker's UDP output had a different schema from what the coordinator read, and its json.dumps(msg, default=float) raised TypeError on any detected object (numpy outline array), which killed the worker thread because only OSError was caught. (2) The frame/arena in the coordinator config did not match the tracker's tag-defined arena; deriving the arena from the tracker's corner tags at startup removes a manual step. (3) 'Delivered' objects must be masked out of detection, or anything put down in view gets re-assigned forever; a boundary band that turns detections into obstacles fixes it. (4) Retry policy: count consecutive misses per robot, not per object id, because a missed object often drops out of tracking and returns with a new id. Also useful: a pre-flight tool that listens for a few seconds, pokes each robot with a zero-velocity command to elicit telemetry, and prints PROBLEM/WARNING lines.