Camera to
silicon.
A custom eight-class YOLO26m detector trained locally, exported to ONNX and deployed through Qualcomm QNN on the Snapdragon Hexagon HTP in a Galaxy S24 Ultra.
Move through the electrical field.
Move your pointer across the module. The live field traces the path from the Snapdragon compute mark to your position.
Official Snapdragon wordmark used for platform identification. Snapdragon and Qualcomm are trademarks of Qualcomm Incorporated; this is an independent project.
Measured on the held-out test.
The public baseline uses the final 306-image held-out test, not an earlier validation or smoke run. The newer RoverCam26m checkpoint remains documented separately.
One model. Two deployment paths.
The raw ONNX build keeps the work portable and testable. The QNN build takes the same perception problem onto the phone’s dedicated accelerator.
YOLO26m trains from scratch at 736 × 736 on a custom Cars4Mars dataset.
ONNX creates a portable graph with 4 box channels and 8 class scores.
Qualcomm QNN routes the Android session to the Hexagon HTP/NPU backend.
CameraX, box decoding and class-aware NMS deliver rover-ready detections.
Eight objects. Competition context.
The model sees the objects needed for target localisation and the rover’s ordered autonomous sequence.
- Hammer
- Tennis ball
- Traffic cone
- Black balloon
- Blue balloon
- Pink balloon
- White balloon
- Yellow balloon
Inspect the system.
The repository includes the browser model, Android source, QNN deployment model, training notes, file hashes and the device-discovery fix.