PhotonVision on SystemCore

[kevinwang]

Just sharing my experience getting PhotonVision running on SystemCore.

Installation is very simple, just copy the PhotonVision JAR onto the device and create a systemd service to run it as root. With this method, most of the functionality works and the web UI is reachable at robot.local:5800. The team number in the configuration can simply be set to localhost to connect to the NetworkTables server.

However, the camera calibration function relies on an external dependency (mrcal) that itself seems to have a lot of transitive dependencies. It seems like it might be difficult to bundle that all into an .ipk package, but I think it would be very feasible to instead put PhotonVision and mrcal into a Debian-based Docker container (Debian because the easiest installation method for mrcal is using apt).

With one 640x480 USB camera at 30fps with a 2D AprilTag pipeline, the CPU usage on the device went from 25-30% without PhotonVision to 65-70% with PhotonVision. Unfortunately I didn't have a higher-resolution camera like an Arducam to test with.

[gerth2]

the device went from 25-30% without PhotonVision to 65-70% with PhotonVision.

While I don't want to speak for the whole photonvision dev team - this is the concerning part. Without a viable strategy to limit CPU load, PV will take down system resources to the point robot code doesn't execute.

The initial recommendation will likely be "Don't install PV on your Systemcore"

[kevinwang]

That's an added benefit of running PV in Docker then -- you get resource constraints (e.g. CPU/memory limits) for free!

[kevinwang]

Here's a working Dockerfile for PhotonVision, it builds a 523 MB image:

FROM debian:bookworm-slim

WORKDIR /opt/photonvision

RUN <<EOF
apt-get update
apt-get install -y --no-install-recommends openjdk-17-jdk-headless libatomic1 libmrcal-dev
rm -rf /var/lib/apt/lists/*
EOF

ADD https://github.com/PhotonVision/photonvision/releases/download/v2025.3.2/photonvision-v2025.3.2-linuxarm64.jar /opt/photonvision/photonvision.jar

CMD ["java", "-jar", "/opt/photonvision/photonvision.jar"]

Build image with:

docker build -t photonvision .

And the run command should look something like:

docker run -d --restart=unless-stopped --name photonvision -p 5800:5800 --cpus 1 --privileged -v /dev:/dev -v photonvision_config:/opt/photonvision/photonvision_config photonvision

[gerth2]

Thanks, good to know there is a path - I'm not familiar enough with docker to comment specifically, but looks like something to look into.

Note there are more ports than just 5800 in use.

[kevinwang]

Good point, it's probably better then to use --network=host instead of publishing individual ports, seeing as the exact list of ports varies based the number of cameras. This way the team number in the configuration can also remain as localhost instead of the more opaque host.docker.internal for container -> host communication.