Thousands of postcards: trials and tribulations

My postcard robot project is an excellent example of the difference between a prototype and a real product. After getting the robot writer working and making the video, I set out to have it produce thousands of cards. It was exciting to finally see my robot in action, that is until it ran into someone with this address:

北京第二外国语学院, 1, 白杨大道

Given that my robot was trained to write latin characters, this was a real problem. Although this was only the beginning of the challenges between me and sending out the cards. When you run a system like this thousands of times, with all kinds of input, you start to uncover a wide range of issues that you never saw before.

With the daydream of the robot fading fast in my mind I started to plow through issues as they came up, beginning with things that the robot couldn't write. You wouldn't believe how many robot-unprintable characters exist in your names and addresses. I ended up devising a complicated scheme where I taught the robot certain common characters, and then flag and sort out cards that could not be printed to be processed later. The hard cards ended up going through a range of mitigations from manual generation to printing labels out on a Dymo for the handful of hardest cases.

More worrying than the software issues was that the robot found several ways to destroy plotters.

This would happen when, for one reason or another, the plotter would not go to the home position. The robot would assume the coast is clear and attempt to pickup a card with the plotter gantry sitting between it and its target.  Long story short, this destroys the plotter.

Knocking out all these issues turned into a very solid integration hell session. Every time I thought I had squashed the last issue, several new ones would pop up.

I was particularly worried by all the failure modes because the robot is kind of dumb. It will happily carry out the motions of writing a card then mark the card as produced in the database - even if something went wrong. The idea of manually going through every card after something bad happened to cross reference the database and figure out what wasn't printed led me to building a camera system. Thankfully I didn't have to use it, but even if things REALLY went off the rails, this would have a log of every card before and after writing to make it easy to go back through time and figure out  what happened and what had to be redone.

There was one issue that I couldn't solve despite a lot of trying. From time to time the handwriting generation would produce squiggly gibberish. I would ask it to write something like "I hope that you enjoy this postcard" and this is what would come out

This would happen with no discernable rhyme or reason. I spent days trying to detect and deal with this, but ultimately I had to begrudgingly admit that I already spent too much time on this and threw in the towel. I had the software generate a message that I liked, which I used across all the cards. I would have really liked to have gotten this part working, but I just had to accept that the cards are robot written, generated, and have a personalized salutation so I'll give it a C+.

They say the last 10% takes as long as the first 90% and I think they're right. This turned into essentially a whole new project, though sadly without a video to go with it. If I would have realized the number of issues that would come up I would have filmed the process.

Eventually the robot did run out of issues and it was really cool seeing that huge stack of cards accumulate next to the robot. The one thing that the robot didn't do is stamp the cards, but my wife was happy to do that. She says she enjoys it. Crazy.

If you put your address into patreon and haven't gotten your card yet, hopefully the mail is just slow. I am still hearing reports of folks getting them over a month after sending some out...

Thank you all for the support!

- Shane