Foundation-Model Training

The model trains by rebuilding what it cannot see. A self-supervised masked-reconstruction objective hides parts of every window under six structured mask types — from scattered random joints and whole temporal blocks to anatomical graph regions and combined spatiotemporal blocks — and asks the network to reconstruct the original clean 3D coordinates. Each mask family forces a different competence: temporal blocks demand motion continuation, graph regions demand anatomical inference, combined blocks demand both at once.

Two properties are earned at once. Per-joint precision: reconstructions stay accurate joint by joint. And an organized embedding space: similar movement lands close together regardless of camera angle, mirroring, or body size. Zero annotations — no labels, no event tags, no human grading — are used anywhere in training.

A model of the body must not invent impossible bodies — bones that stretch, joints that teleport. Throughout training, outputs are held to anatomical plausibility: bone lengths stay consistent within a window, and velocity and acceleration profiles stay smooth. The result is not a physics engine; it is a soft anatomical contract that keeps every reconstruction a body.

A 2D skeleton is a 3D skeleton seen through a camera — a projection that discards depth. To this model, depth is simply one more thing that can be missing: because it has learned how bodies move in three dimensions, it restores the lost dimension the same way it restores an occluded joint. The same forward pass that denoises and fills gaps also lifts.

The current checkpoint comes from a proof-of-concept run: 8×H100 for 14 days — roughly 2,700 GPU-hours — trained end to end with zero labels. It is a research checkpoint, not a product model; its measured behavior is documented in the evaluation suite.

Self-supervision converts raw footage into training signal, so capability scales with data, not annotation budgets. What the model learns from is movement data at scale — a massive and growing corpus of movement video, converted into skeleton data. The ladder below is the honest map from corpus scale to what the model can support — as the training corpus grows by orders of magnitude, harder capabilities become possible.