Text reading, templates and object detection have different jobs
OCR reads visible text. Template matching looks for visual similarity to a saved reference. An object detector predicts where a category of object appears. Each can be useful, but a readable label is not an animal identity and a detected object is not necessarily the intended errand target.
Targeted crops reduce irrelevant screen content. They also define a boundary: a detector trained on the exposed City Event map area is not automatically qualified for every screen, resolution or interface arrangement.
Recognising animals and resource objects
VRA’s internal visual experiment covers deer and sheep alongside berries, ore, timber and material piles. Reviewed examples include different poses and lighting; complete capture runs and a separate farm are kept outside training for evaluation.
The first model work established an evaluation and local execution pipeline, with controlled animal trials and offline resource checks. Extra detections and missed objects remain important. The expanded model has not been released as an unattended resource-clicking capability.
Why a moving animal needs a current observation
A gold guide marker can indicate where an errand was introduced, while the animal itself moves away. Following the old marker indefinitely can send input to empty ground. A useful perception system must retain the target’s identity while updating its position.
Multiple observations and body tracking can help, but ambiguous targets, stale frames and uncertain tracking should constrain action. A model score is not a guarantee of correctness, and periodic recognition checks are not proof that a click will land.
Training and live automation are separate
Offline training changes the model using labelled examples. Live inference uses a fixed model to produce observations. Simply running a farm does not currently retrain VRA or grant a model new permissions.
The customer-contribution plan is optional and separate from core automation. It requires explicit consent and release safeguards; current public pages are not an active request for private screenshots. A later experimental model-control direction must pass internal comparisons before release.
What a useful visual-automation result should prove
A strong demonstration shows the relevant starting state, current target, attempted action and visible result. Model export checks establish that two runtimes produce equivalent predictions; they do not establish perfect detection, fleet reliability or safe account use.
For a gathering task, a correct visual box is only one step. The workflow still has to respect the configuration, handle existing activity, send an authorised input and observe the expected effect. This is why VRA publishes development milestones with their qualification boundaries.