Research status
AutismAI Lab builds machine-learning systems that turn ordinary video and everyday tablet use into individual-level evidence for autism spectrum disorder — and spends just as much effort checking whether those measurements can be trusted.
Below is where that work stands, in plain terms: what is already out for peer review, what is finished and about to go out, and what is still being written. Every claim on this site also traces back to a source — the peer-reviewed references those tools build on are listed further down this page.
Stage 1
Sent to a journal or conference and currently being evaluated by outside reviewers.
Tests whether an interpretable ensemble trained on markerless skeleton sequences from everyday imitation tasks can distinguish autistic from neurotypical movement, with attention analysis designed to point back at which movement phase and which joints drove each decision.
A model trained only on neurotypical movement, with no autism examples at all, scoring how far an individual’s movement deviates using reconstruction error rather than a per-feature z-score.
Audits whether a widely used stereotypical-motor-movement benchmark rewards models for pose-tracking quality rather than for behaviour, and proposes a frozen evaluation protocol that scores every model on both a standard and an artifact-controlled split, reporting the gap between them as a first-class metric.
A human-in-the-loop study with caregivers and licensed clinicians testing whether separating “is this the right object” from “is this content okay for my child” actually helps catch AI errors before they reach a child.
A within-subjects caregiver study comparing a review queue that surfaces the model’s least-confident extractions first, with a one-line reason attached, against a plain in-order queue — measuring whether prioritization changes how many real errors are caught, and at what point it stops helping.
Stage 2
Research and writing are done; in final author review before going out to a venue.
The verification gate measured against human judgment on public benchmarks, plus a calibrated decision rule for when the gate should approve automatically versus route to a caregiver — reported with the review-burden cost of that guarantee, not only its benefit.
Stage 3
Still being written or awaiting final study data before a venue is chosen.
Age- and sex-conditioned Gaussian-process reference models of neurotypical movement, producing a per-individual deviation score across the full kinematic feature set rather than a group-level comparison.
An early-stage extension of deviation scoring from individual feature values to entire movement sequences: a masked, condition-aware model predicts hidden parts of a sequence from its visible context and scores deviation on what was hidden.
The two programs
Turning ordinary video of everyday movement into objective, individual-level measurement — and holding that measurement to the same rigor a clinical instrument would require.
Volitiq methodology →A home-learning system for minimally verbal autistic children, where every AI-generated piece of content passes a caregiver checkpoint before it reaches a child.
Grovalin methodology →Two of the manuscripts above exist specifically to keep that work honest — the shortcut-robust benchmark audit on the Volitiq side and the verification-gate audit on Grovalin's — because a tool that looks accurate on the wrong evaluation is not evidence-grounded, however good the number sounds. The longer write-up of each research line is on the research in detail page.
Research library
Every claim on this site traces back to a source. These are the peer-reviewed references the two programs build on, each linked to its DOI or arXiv record.
Assistive learning for children with autism, built on evidence-based intervention frameworks and an open-vocabulary computer-vision pipeline.
Key external references
Objective, browser-based motor assessment grounded in sensorimotor-control research and markerless pose estimation.
Key external references
Our standard
External references are peer-reviewed and linked to their DOI or arXiv record. The lab's own manuscripts are listed by status and venue only, described by what each study tested rather than what it found.
We publish a specific result only once it is reproducible. That is why no accuracy, AUC, effect size, or usability score from a manuscript still under review appears anywhere on this site. Where a reproducible result does exist today — Grovalin's macro-F1 on the public COCO-2017 and Open Images V7 benchmarks, which anyone can re-run — it is stated in full on the relevant science page.
We work with researchers and clinicians on shared studies, datasets, and manuscripts. If our methods are useful to your work, we would like to hear from you.