MotionBricks: give a character the poses, let the model find the movement

The demonstration keeps the movement connected

The project’s uncut Unreal Engine 5 showcase runs for two minutes and forty seconds. It combines navigation styles and object interactions within one environment. The accompanying examples isolate tasks such as picking up a sword, jumping over a bench and sitting. They make it possible to inspect the approach and exit around an action, rather than seeing only its most attractive pose.

Those transitions are the subject of the research. In a conventional animation system, clips and rules determine how a character moves from one state to another. MotionBricks supplies a learned generator that can connect specified states. The application still tells it where to go and what pose to reach.

Sources: NVIDIA and research collaborators [1] · Wang et al. / ACM TOG, SIGGRAPH 2026 [2]

A short horizon, updated repeatedly

The runtime receives four recent poses and between one and four target poses. It generates a short continuation, sampled at 30 Hz, that can become the context for the next request. Continuous animation is assembled through this cycle of generating, playing and replanning. A changed target can influence the next continuation.

Three learned components divide the work. A root module predicts timing and the character’s overall trajectory. A pose module predicts compact motion tokens. A decoder turns those tokens into continuous motion features. In practical terms, the system has to decide both where the body travels and how its joints move along the way.

The output is skeletal motion that an application can use on a character. It does not contain the rendered costume, lighting, environment or camera work seen in the demonstration. This matters when evaluating its role in games or cinematics: MotionBricks operates on the performance layer of an existing 3D scene.

Sources: NVlabs [5] · NVlabs [3]

The useful unit is an interaction

A Smart Primitive stores target poses in relation to an object, together with an interaction region and an ordered sequence. A pivot can adapt the sequence to the character’s approach. The released editor’s avoidance example places an entry pose before a bar and a second pose beneath it. The model generates the movement between them from the character’s current motion.

The change is in what gets authored. An interaction can be described through the places and poses that matter, while the model supplies a plausible route through them. Moving an object does not necessarily require a separately recorded approach for every direction. That is the intended reuse; it is not a guarantee that any target arrangement will look natural.

Sources: NVlabs [6] · NVlabs [7]

What the download actually contains

The standalone package includes pretrained SOMA and Unitree G1 models, ONNX inference, an Inbetweener, Scene Preview, a Smart Primitive Editor and a Scene Editor. The runtime uses Python with Newton, Warp and TensorRT. The repository says its demos were tested on Windows and Linux with an RTX 4090.

NVIDIA explicitly says the Unreal Engine integration is not included. The Python implementation carries over the research’s core ideas for developers to inspect and integrate. Nor is the bundled editor a finished production tool: its documentation warns that it has no undo or redo. That is a practical constraint for anyone attracted by the polished UE5 video.

The SOMA card identifies 313,526 training clips from Bones RigPlay 1, prepared for a 30-joint skeleton. The model’s inputs are motion features; text prompts and scene geometry are outside its interface. A new character rig needs a suitable representation or retargeting. Obstacles and target placement remain the application’s responsibility.

Sources: NVlabs [3] · NVlabs [4] · NVIDIA Developer [9] · NVlabs [8]

15,000 FPS is a motion-generation result

The paper reports 15,000 motion frames per second and 2 ms inference latency on an RTX 5090. These figures describe the neural motion system, not the frame rate of the rendered Unreal scene. They do not include everything a game must do to display a frame, and they should not be read as a measured crowd capacity.

There is another boundary between the paper and the download. The SOMA model card says the paper evaluates its own models and datasets, not these released ONNX exports. It does not provide a complete end-to-end quality benchmark for the exported models. The paper’s speed headline should therefore remain a research result, not a promise for a particular installation.

The authors also state a substantial boundary: generated kinematic motion can be physically implausible. The robot demonstration therefore uses a separate tracking controller. For animation, the corresponding concerns include self-intersections, contact quality and unfamiliar target poses. The current model card does not promise equivalent results across skeletons, precision settings and inference backends.

Sources: Wang et al. / ACM TOG, SIGGRAPH 2026 [2] · NVlabs [8]

Control moves closer to the scene

MotionBricks is compelling because its controls describe visible actions: approach here, pass through this pose, reach this contact. The release gives developers enough of that system to examine the idea beyond a trailer. The next useful comparison is specific: how reliably does a reusable interaction survive a different approach, a moved object or a new character? That is where a convincing animation method earns a place in a project.