Why small errors cause big losses on a packing line
A small error on a packing line becomes a big loss when it stops the line, because a stop costs more output than a missed pick. TOS Robotics learned this on a bell pepper line in the Netherlands, where a stop of our robots also sent a person to reset another machine. We rebuilt our software to retry and recover first, and to stop where a person has to act.
First setup
Our first control software came from the lab, where an error means stop and let someone look. A V-shaped belt brings the peppers to two robots, and after a run of failed picks the software cleared that belt or shut the whole system down.
Why a stop costs so much
The two robots fill the same pockets of the flowpacker infeed, the compartments that carry the peppers into the packing machine. If one robot stops, the other leaves the packs half filled, so a fault on one robot now stops both on purpose.
When our robots stopped, the upstream singulator, the machine that puts the peppers in single file, went into an error that someone had to clear at the machine. Each stop we caused meant someone had to walk over.
We split the whole line’s lost output per day into three parts: the line standing still, pockets the robots never reached, and failed picks. On one bad day the line standing still cost about seven times more packs than both robot losses together. A long stop hardly shows in a robot’s success rate.
What we changed
We rebuilt the reactions in one order: retry, recover, warn, and only then stop. A recovery is a short belt reverse and an air burst, so the pepper lies differently for the next try.
| Event | First reaction | Reaction now |
|---|---|---|
| A failed pick | Clear the belt after 3, shut down after 10 | Recover and retry, clear the belt at the 5th failure in a row, warn, then stop |
| A short flicker from the singulator | Both robots stopped | The signal must last half a second first |
| The PC freezing about 3 s on disk writes | Run stopped after 250 ms without camera images | Belt stops after 250 ms, the run only after 3 s |
Over two weeks, each self-stop after repeated failures was followed by good picks once the line restarted, so we raised the limit. A belt clear-out costs several seconds, and the pick after one failed no less often than a plain retry. In the recorded data, a clear-out at the 5th failure instead of the 3rd would have removed most clear-outs.
A morning of repeated stops logged only as “Press Reset or close the door” led us to give each singulator alarm its own code and instruction, see fault screens that operators actually see.
Results
The robots on this line run at 99.3% uptime. We have no before and after figure that isolates the effect of these changes.
Small percentages also added up. A few percent of the pockets were missed by the robot’s belt tracking, a few percent of feeds sent the belt two stop commands, and after a recovery a pepper lay badly many times more often than after a normal feed. When the belt was sped up, the two robots could no longer reach every pocket, so cycle time now limits output next to accuracy.
Questions and answers
Should a robot cell ever stop itself?
Yes, when a person has to act, for example when a door is open. For other events we retry, recover and warn first.
How do you keep short glitches from stopping the line?
A signal must stay wrong for long enough before the cell reacts. In our data nearly all real stops of the singulator were still active several seconds later, so waiting half a second delays a real stop only a little.