The science behind Grovalin.

Autism spectrum disorder affects about 1 in 36 children in the US.^{[Maenner et al., 2023]} Among the most documented and most stubborn barriers in early intervention is concept generalization: a child learns a label tied to a single image and cannot transfer it. They identify their cup but not a different cup; they learn "dog" from one book and miss the dog across the street.

This is not a motivation problem or a memory problem — it is a transfer problem, and it compounds. Vocabulary that does not generalise does not become functional communication.^{[Tager-Flusberg & Kasari, 2013]} Grovalin's design goal is narrow and specific: make the same concept appear in many forms, so the category — not the photograph — is what the child learns.

The intervention scaffolding is grounded in established evidence-based practices for autism, rather than invented from scratch.^{[Wong et al., 2015]}

A caregiver records short videos of everyday objects. Grovalin turns them into a personal learning library through a four-stage pipeline that needs no fixed list of object classes — essential, because a child's world contains objects no pre-trained detector was trained to name.

Detect. Grounding DINO performs open-vocabulary detection, locating objects from a free-text prompt rather than a closed taxonomy.^{[Liu et al., 2023]} Segment. SAM 2 cuts each object cleanly from its background into a transparent asset.^{[Ravi et al., 2024]} Verify. An OpenCLIP vision-language check confirms the crop matches the intended label and flags truncated or ambiguous objects.^{[Radford et al., 2021]} Vary. Colour variations are produced with HSV image processing and structural variations with a rectified-flow diffusion model.^{[Esser et al., 2024]}

Single-viewpoint authenticity. Grovalin deliberately does not synthesise novel 3D views or pull canonical product photos from the internet. The child sees the object as it sits in their home — partial views and odd angles included — because that is how they will encounter it in life. Familiarity is the point, not a limitation.

To avoid the circularity of scoring a model against its own labels, the detection-and-verification pipeline is evaluated on held-out public benchmarks — COCO-2017 val and Open Images V7 val — rather than on Grovalin's own data.

At a detection confidence threshold of 0.30, the pipeline reaches a macro-F1 of 0.734 on COCO and 0.550 on Open Images, with high recall on both. These figures are reproducible from the evaluation set and are reported in the technical manuscript below.