We ran into an issue where we didn’t pass inspection due to our frame wasn’t isolated. After checking we found the source of our problem was the Google Coral we added last week. It appears that the heat sink on the Coral is not isolated from the frame. If you are using a Google Coral make sure to not mount it to conductive material.
I haven’t seen a Coral used but from my understanding it’s an SBC like a Jetson nano or a raspberry pi correct? That should be grounded through its power supply, but also all of those SBCs have solder points on the bottom. They should always be mounted with standoffs. Otherwise you are shorting all of the points on the bottom of the circuit board. There’s lots of 3d printed cases that would even work to just act as an electrically isolated mounting plate.
Out of curiosity have you used any other SBC for comparison and if you don’t mind me asking how you are using the Coral this year? I’m assuming object recognition or some other ML?
The coral is enclosed so there aren’t any exposed solder points which is why we mounted it as we did. Live and learn!
We are using it to recognize cones vs cubes and adjusting our pick-up based on which object. Really happy with it, we have only had it for a week and it was very easy to train and integrate it.
Follow Up question if it’s not asking too much. What system are you using underneath for your ML? I’m not super familiar with Corals limits but I’m very familiar with the Jetson Nanos and pis limitations here and using YoloV3/v4 with custom trained images. I’m assuming Coral is using something similar if not exactly that same type of setup and training process? Was this difficult for your team to do? You said it was very easy, I just don’t know know your definition of easy versus what my students would be. I don’t think training is hard but when I did it about a year ago it wasn’t the most user friendly process and training was very tedioua
Coral is just an accelerator coprocessor. I have used it to accelerate ML models built using both Axon and Teachable Machines - export as an Edge-optimized .tflite model. There is a python wrapper that makes it pretty easy to use.
To add Coral support to a WPILibPi (or other Raspberry Pi coprocessor):
echo "deb https://packages.cloud.google.com/apt coral-edgetpu-stable main" | sudo tee /etc/apt/sources.list.d/coral-edgetpu.list curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo apt-key add - sudo apt-get update sudo apt-get install python3-pycoral # Run Coral TPU at normal clock rate: sudo apt-get install libedgetpu1-std # --or-- # Overclock Coral TPU at double clock rate: sudo apt-get install libedgetpu1-max
Here’s a quickie python script and crude classifier I threw together several weeks back.
tensorFlowServer.zip (563.2 KB)