results

Reliability

The computer vision node was able to perform very reliably, capable of locating and identifying the card number and suit with the deck it was trained on, achieving a >99% mean Average Precision (mAP) on a validation set of synthetic images it wasn't trained on.

There was some unreliability in the exact coordinates that the vacuum gripper made contact with the card during pick-and-place, and the paths planned by MoveIt were sometimes unnatural, but this inaccuracy did not prevent successful manipulation of the objects. However, more fine-tuning would be required if we attempt to expand object manipulation beyond cards into poker chips, etc.