Concept guide · SIGGRAPH 2026

Smart primitives: the second half of MotionBricks' name

The paper title pairs a “modular latent generative model” with “smart primitives.” The backbone gets the headlines — 15,000 FPS, 350,000 clips — but the primitives are what a user or engineer actually touches. This page explains each family as the official material describes it.

Locomotion: three inputs, no per-task training

The locomotion primitive takes continuous commands — velocity, heading, and a style signal — and the backbone renders them into full-body motion. There is no clip per combination: you are not selecting “walk-forward-zombie,” you are setting dials and the model synthesizes. That is where zero-shot composition shows up in practice; adding a movement flavor does not require training a dedicated model for it.

The interactive demo makes the idea tangible. WASD sets direction and speed relative to the camera; each style key layers a character onto the same locomotion stream — 11 styles in the shipped build:

KeyStyle
VSlow walk
ZHand crawling
XWalk boxing
BElbow crawling
RStealth walk
TInjured walk
CWalk stealth (crouched)
EHappy dance walk
FZombie walk
GGun walk
QScared walk
Zombie-walk locomotion style generated live from velocity, heading and style inputs
Zombie walk — same locomotion primitive, different style input.
Crouched strafing locomotion blending direction and style without authored transitions
Crouch strafing — blended direction and style, no authored transition.

Objects: proxy keyframes the backbone fills in

Object interaction works differently. The author supplies proxy keyframes — a rough specification of the interaction, like where a hand should meet a sword or where a bench sits relative to a jump — and the backbone synthesizes the surrounding motion: the approach, the contact, the follow-through. Because those segments are generated rather than played back, repeated runs vary naturally instead of replaying one baked take. The official GIF gallery shows the family: pick up sword, jump over bench, sitting, falling.

Worth stating plainly, since the UE5 page gets asked this weekly: there is no collision solver underneath. Contact behavior is learned into the motion, which is exactly why the material claims no foot-locking, no blending, no collision detection and no hand-authored transitions. If your project needs simulated physics response, this is a motion generator, not a physics engine.

Setup: authoring end-to-end inside UE5

The third leg in the official material is the authoring workflow: NVIDIA's Smart Primitive Setup section describes setting primitives up inside Unreal Engine 5 itself — placing proxy keyframes on scene objects and wiring the behavior without leaving the editor, which is how the 2:40 one-take demo was assembled. No public UE5 plugin ships in the preview, so today you can read about it on the project page and run the model layer via the Python setup guide.

The state-machine question (our opinion, labeled)

Opinion, not official: read together, the three families sketch an alternative to the animation blueprint state machine. Instead of authoring states and transition rules between every pair, you expose a low-dimensional command surface and let one trained model occupy the space — the “blending graph” moves from the game project into the checkpoint. That reframing is ours; the paper's own pitch is scalability and real-time control. Run the demo, press F into a zombie walk while strafing, and judge the analogy yourself. The controls are listed on the G1 demo page, and the home page maps the rest of the framework.