The camera was holding up the steering. I split the system, and it placed 4th nationally.

The problem
The first prototype ran vision and motor control together. When object detection got busy, the controls had to wait. A vehicle that hesitates every time the camera thinks isn't much of a vehicle.
What I did
I split camera capture, YOLO inference, video streaming and motor commands into separate multithreaded pipelines and Docker services on a Raspberry Pi 5. Then I built a native Android control station that brings live video, driving and telemetry into one screen. If the Wi-Fi drops, the vehicle brings up its own fallback hotspot.
The result
A 30 FPS camera feed with on-device detection, live video and telemetry, and 4th place plus a Special Award at the 10th Galaksija Cup National STEM Championship. The code is open source under the MIT license.

Check it yourself