The demo failed on stage. Four days later I knew why

On September 29 we presented Gemm at the Parque de la Innovación Community Meetup. We brought the G1, a rehearsed talk and the demo I cared about most: someone in the headset moving their arms and the robot following them in front of everyone.

The day went wrong early. The robot's own network, which had worked perfectly days before, stopped coming up, and there was no way into the robot over wifi or cable. The cable turned out to be faulty. We ended up plugging a monitor, keyboard and mouse into the robot to get in directly.

We got access back, recorded the teleop working at ORT just in case —it's the video in the Gemm case study— and walked the few blocks to the Parque with the robot and everything we needed to set up our own network.

On stage

We had prepared ten minutes; they told us we had three; we got seven. We set up the network, tested, and teleop worked. While the previous group was presenting we turned it on again, and this time it wouldn't. The robot stood next to the stage and we presented anyway, with Eitan's dashboard video running on recorded data.

I was left with one question: if it worked at ORT and worked when we arrived, why did it break right then?

What I found afterwards

1. The robot's network didn't always start

The next day I found why: when starting up, the robot's own network competed with its own wifi connection, and sometimes lost. I rebuilt it differently and added a way out: if nobody connects for a while, the robot goes back to the normal wifi by itself. In the tests after that it ran without drops.

2. The video died as soon as the headset saw a hand

On October 2 I built a tool to measure, from the headset, what was arriving and what wasn't. One test made it obvious: video worked fine until the headset detected a hand. At that moment the library tried to download something from the internet, and the robot's network has no internet. It failed, and took the video down with it. I fixed it by keeping everything needed on the robot itself. After that: almost a minute in VR with hands, without a single drop.

I can't say for sure which of the two killed the demo on stage. I do know both only showed up under the event's exact conditions: the robot's own network, no internet. At ORT, with internet, neither was visible.

October 3

Full teleop with arms and high-definition video in the headset. Everything worked.

What I take from it

• —Test under the same conditions as the demo, not similar ones. The whole difference was "no internet" versus "internet".
• —When something works only sometimes, measure before changing things. Measuring solved in one test what days of guessing didn't.
• —Bring a recorded plan B. The teleop and dashboard videos let us tell what we couldn't show.