kageroumado/rocuronium-ui-detector

Rocuronium UI Detector

A tiny, fast UI-element box detector for desktop screenshots, exported to Core ML for Apple silicon. It proposes bounding boxes for on-screen controls so a GUI agent can aim at an app whose accessibility tree is empty or lying — icon toolbars included.

Trained for Rocuronium, a macOS UI-automation tool that drives apps without taking the cursor. It is the detector tier of Rocuronium's vision-grounding cascade (accessibility → detector + OCR → VLM).

What it does

• Input: an RGB image, 640×640.
• Output: confidence [N, 1] + coordinates [N, 4] — non-max-suppressed boxes in the standard Core ML / Vision VNRecognizedObjectObservation format.
• One class, UIElement. The model answers "a control is here", not "this is a button vs a checkbox". Role and label come from OCR text inside the box (and, in Rocuronium, a later VLM caption pass). This is deliberate — see Taxonomy below.

Why single-class

GroundCUA's category field is 8 coarse buckets with ~33% of boxes empty, and it cross-cuts visual type — there is no clean signal to learn Toggle / Slider / Checkbox / Link / Icon from. A box-proposal detector paired with OCR for the label is both what the data supports and what the downstream consumer actually uses, so the model commits to proposing boxes well rather than guessing a role badly. A 3-class variant (Text / Interactable / Icon) is a documented follow-up.

Metrics

Trained on a stratified ~13K-image subset of GroundCUA spanning all 87 apps, yolo11n, 640px, 100 epochs on Apple MPS.

Metric	Value
mAP@50	0.881
mAP@50-95	0.479
precision	0.864
recall	0.851

Core ML package: 5.4 MB. Latency: ~8 ms/image mean on .cpuAndNeuralEngine (Apple silicon), 5.7 ms min.

Usage (Core ML / Vision)

import Vision
import CoreML

let model = try VNCoreMLModel(for: MLModel(contentsOf: compiledModelURL))
let request = VNCoreMLRequest(model: model)
request.imageCropAndScaleOption = .scaleFill
try VNImageRequestHandler(cgImage: screenshot).perform([request])
let boxes = (request.results as? [VNRecognizedObjectObservation] ?? [])
    .filter { $0.confidence >= 0.25 }   // normalized boundingBox, bottom-left origin

Training data & license

Trained on ServiceNow/GroundCUA (Apache-2.0). The weights are released under Apache-2.0 — unlike existing GroundCUA-trained detectors, which inherit Ultralytics' AGPL-3.0 and cannot ship in a closed app. Ultralytics tooling was used only to train; the exported weights carry no Ultralytics code.

Limitations

• A pre-filter, not a labeler. No role, no text — pair it with OCR.
• Trained at 640px; very small or very dense controls may merge or drop.
• Desktop UI only (GroundCUA is macOS/Windows/Linux desktop apps); not tuned for web or mobile layouts.