Game animation has been stuck in the same loop for decades: animators hand-craft transitions, tag motion clips, build state machines, and wire up animation graphs. The result is brittle, expensive, and doesn't scale. NVIDIA Research's new MotionBricks framework, accepted at ACM SIGGRAPH 2026, makes a credible case that all of that scaffolding can be replaced by a single neural model running in real time.

One model, everything it needs to know

MotionBricks is a real-time generative framework that combines a large-scale latent backbone with intuitive "smart primitives," delivering high-quality, zero-shot motion synthesis at 15,000 FPS. Zero-shot here means the model handles new tasks without any fine-tuning or task-specific data tagging -- you just describe what you want and it figures out the motion.

The single neural backbone covers over 350,000 motion skills, trained on a corpus of 350,000 production-grade mocap clips from real human actors and actresses. That dataset, called BONES-SEED, is also open-sourced as part of the release.

The two-part architecture

The paper identifies two concrete problems that have kept generative motion research out of production: real-time scalability and integration. Existing generative models degrade badly under real-time constraints, and text- or tag-driven interfaces don't give artists the fine-grained control they need. MotionBricks attacks both problems simultaneously.

• Modular latent generative backbone: A VQ-VAE (a neural encoder that compresses motion into a discrete codebook, like tokens for motion) sits at the core. The repo ships three separately trainable components --