Google EdgeTPU
An AI based sorting machine built to show off Google’s new Edge TPU development board
Contribution
As Digital lead on the Agency side I led the ideation process that resulted in the idea to build a computer vision based sorting machine that could show off the speed and advantage of Edge computing that could run on a small inexpensive Coral Development Board.
Key Results
I was able to lean on my background in interactive experiences and physical computing to lead a team that build an MVP in a very short time period.
We made several improvements to the installation over a year as the installation travelled to 4 shows based on user feedback, client feedback and technical changes.
Process
MVP (V1)
For the MVP we built a low budget version of the sorting machine where we used a low power motor to run an aluminium chute that sent plastic pucks into the right slot. We also had a simple UI on a monitor that showed We were looking to prove a few crucial details:
Optimize all the physical aspects of the installation
Show the difference between edge computing vs cloud computing based computer vision
Demonstrate the value of edge computing in a realistic scenario
This version was shown at Mobile World Congress in Late October 2018.
V2
The MVP was deemed a little sluggish so we were able to make an upgrade to a larger stepper motor with a more robust gearing setup to make it feel more snappy. This gave us the following benefits:
Allow hardware to catch up with the TPU’s processing speed
Bring the overall installation to a more real world feel which was something the client wanted to achieve
This version was shown at the annual Google Cloud event in late November 2018.
Final Version (V3)
With two successful runs, we were finally able to build a case for the most optimized version of hardware and software. Updates on the software side meant the program was running faster so we were able to get a carbon fiber chute for V3. This allowed us to achieve the following:
Create a close to real world scenario and something that looked impressive to attendees at events
Allow hardware to catch up with software speed
This version was displayed at Google I/O 2019 and Google NEXT 2019
