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.

Reconstruction alone yields detail without organization. A multi-view consistency objective adds the global structure: two augmented views of the same window must land near each other in embedding space, with regularization terms preventing collapse — no negative pairs, no labels. Reconstruction provides per-joint precision; consistency organizes the embedding space. Zero annotations are used anywhere in training.

Reconstruction losses alone will happily produce plausible-looking nonsense — bones that stretch, joints that teleport. Physics regularization penalizes exactly that: bone lengths must stay consistent within a window, and velocity and acceleration profiles must 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. The model therefore treats lifting not as a separate task but as reconstruction from a lossy view: during training, windows are projected through random virtual cameras, and the target is always the original 3D sequence. Depth becomes one more thing that can be masked. The same forward pass that denoises and inpaints 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. 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.