MOTEM
booting…
pose
tap to close
motion embedding

SPACE — choose features, then build it

features

Rough space needs no calibration: a presampled corpus of synthetic motions is embedded once (~10 s) into a fixed map — shown as a grey cloud — and your tracking lands on it immediately. Same seed on every device, so rough-space sessions stay comparable. For a space tuned to your movements, calibrate below (decoded replays only make sense after real training). Export downloads the weights; commit them to model/ in the repo and every visitor starts in your space.

CALIBRATE — optional, personal space

Record 1–3 minutes of the movement vocabulary you care about (walk, turn, raise arms, squat, wave, dance…). A motion-VAE and a parametric UMAP head are then trained in your browser. All later sessions are embedded into this one frozen space, so they stay comparable.

architecture
0windows
1 · motion VAE
2 · UMAP fit
3 · parametric projector

COMPARE — overlap / novelty

vs

Tap regions of the map (left tile shows the decoded motion — the VAE run backwards from the latent).

MOTEM

Whole-body motion embedding, live. One tile shows the pose the camera sees; the other maps every movement window into a learned latent space (motion-VAE → parametric UMAP). Record sessions, compare their maps, tap a region to replay the motion decoded from the latent.