A hierarchy: space, then time

The Hierarchical Motion Transformer reads movement the way an analyst does: one body at a time, then across time. Each input is a 100-frame window — four seconds of motion at 25 Hz — over the 21-joint skeleton. Spatial attention layers first resolve how joints relate within a single instant: how the knee loads, how the trunk counter-rotates, how the arms balance a cut. Temporal layers then model how those poses evolve — from sub-second micro-adjustments to the full arc of a sprint.

Factoring space and time keeps the model compact — roughly 11 million parameters — without flattening the structure of the body into a generic token soup. The skeleton stays a skeleton all the way through the network.

T × 21 × 34 s window @ 25 Hzspatialjoint-graph attentiontemporalhierarchical time modeltokens + 384-ddual outputs
One forward pass: from raw joint coordinates to per-joint tokens and a sequence embedding.

Attention that knows the body

Plain transformers treat inputs as an unordered set; bodies are not. Joint-graph attention biases every attention score with a learned term indexed by hop distance — how many bones separate two joints on the skeleton graph. A wrist attends to an elbow across one hop; to the opposite ankle across many. The prior is soft: the model can still discover long-range couplings — arm swing against stride, trunk lean against deceleration — but it starts from anatomy instead of rediscovering it.

The 21-joint, 20-bone topology this bias is computed over is documented in the skeleton reference.

Two outputs, one pass

Every forward pass produces two views of the same window. Per-joint tokens carry local detail for reconstruction work — denoising, inpainting, 2D→3D lifting. A single 384-dimensional embedding summarizes the whole sequence for retrieval, clustering, and probing. Downstream products consume one or both; nothing needs a second model.

How these outputs are trained is covered in foundation-model training; how they are measured, in the evaluation suite.

Model datasheet

Prototype
ParameterValueNote
backboneHierarchical Motion Transformerspatial → temporal factorization
parameters~11Mcompact by design — fast to iterate, cheap to serve
input window100 frames · 4 ssampled at 25 Hz
skeleton21 joints · 20 bonesdocumented in the skeleton reference
attention biashop-distance relative positioncomputed on the skeleton graph
outputsper-joint tokens + 384-d embeddingone forward pass, two views
objectivesmasked reconstruction + multi-view consistency + physics termssee foundation-model training