The science behind the tools, one level deeper.

The hand-crafted kinematic features on the Volitiq science page answer "how did this movement differ, in interpretable terms." A parallel thread asks whether a deep-learning ensemble — trained directly on markerless skeleton sequences rather than on hand-engineered features — can separate autistic from neurotypical movement while staying interpretable enough to point back at which movement phase and which joints drove the decision.

The approach combines three complementary encoders (temporal convolution, spatial-temporal graph convolution, and graph attention) in a late-fusion ensemble, evaluated on video of everyday tool-use imitation tasks from a cohort spanning ages 7–34. The attention analysis is built to surface which movement phases and joints carry whatever discriminating signal exists, so that any result can be checked against the motor-planning differences the feedforward/feedback dissociation line is built on.

The Gaussian-process normative models in the shared statistical layer score a fixed set of kinematic features against a population norm. This thread asks a complementary question: can a model trained only on neurotypical movement — with no autism examples at all — flag atypicality directly from the full spatiotemporal trajectory, using reconstruction error as the deviation signal?

It is an unsupervised counterpart to normative modeling: the deviation score falls out of how well a "typical-movement" model can reconstruct a given clip, rather than out of a per-feature z-score.

A methodological thread, in keeping with the lab's evidence-grounded-methodology principle: before trusting any classifier's accuracy number, check what it is actually keying on. This work audits a widely used public benchmark for stereotypical motor movements, asking whether a classifier given only pose-tracking-quality signals — no movement information at all — can still score on the benchmark's standard test split. To whatever extent it can, part of what looks like "detecting behavior" would in fact be "detecting how well the camera tracked the child."

The manuscript pairs that audit with a frozen, pre-registered evaluation protocol that scores every model on both the standard split and an artifact-controlled split, reporting the gap between them as a first-class metric, plus a compact detector designed from the start to be robust to the artifact.

The Gaussian-process normative modeling method already described under "A foundation, not a feature" has its own dedicated manuscript: age- and sex-conditioned reference models built from neurotypical participants, producing a per-individual deviation z-score across the full kinematic feature set rather than a single group-level comparison. It is the flagship paper for that method as applied to Volitiq's motor data.

An early-stage extension of the same normative-modeling idea from individual feature values to entire movement sequences: rather than scoring a fixed set of features against a population norm, a masked, condition-aware model predicts hidden parts of a movement sequence from its visible context, and scores deviation on what was hidden. The goal is sharper, less leakage-prone deviation scores than a standard autoencoder baseline gives.