We are looking for a new coprocessor that can do the following
Run PhotonVision Apriltag detection at a good fps (15+)
Run our custom ML model (5+ fps)
Have a bit of compute left to do other random tasks like record a camera if needed
From some of the research I did the Orange Pi 5 looks like it would work however I can’t find much on running ML models on them. We would probably be buying 4 of these, we were also looking at getting a Jetson Orin Nano but that seems a bit overkill for FRC and I don’t know how well multiple PhotonVision instances would run.
We are just using our old programming laptop. Built-in battery with an i5-6200u.
With barebones Linux and PhotonVision, we get around 20fps on 2 cameras simultaneously or 40+ on one.
If you have a laptop you can use, verify that it is under the $600 limit.
We already had the old laptop and it worked perfectly.
It definitely raised some eyebrows but got the job done.
Some reasons we went with that instead of buying a new mini PC were
You could buy a new version of it for under 600$, so it doesn’t conflict with that rule.
The battery is also enclosed and part of a system, so it doesn’t break that rule either.
(Laptop batteries are actually listed as an example of something that doesn’t break the rule)
All of the components are also rated for impacts because they are meant to be carried around.
(the laptop is mounted on the left support for the center pillar)
We use an old model of ThinkPad (not sure which specifically), and in one of our later matches, it wasn’t fastened properly and fell out of the robot. It was run over by several other robots but still functions perfectly with minor cosmetic damage.
It’s definitely something we are gonna continue using.
You don’t have very hard to achieve requirements for your vision processing. The thread below may help you. You can definitely make use of a RPiRaspberry Pi 4 or if you want to go much faster you can get an orange pi 5 too. This thread might help some the results are not up to date for next year though (next year will use 36h11 Apriltags).
You can most easily do this with a Limelight 3. However, if you want to have some fun, I’llLimelight, an integrated vision coprocessor fund your purchase of an Orange Pi 5 if you integrated the ML stuff into Photonvision. The Opi5 has around 2x the performance of a LL3 with a single camera, and probably significantly more (not tested) with multiple cameras.
