NVIDIA's MotionBricks runs 15,000 FPS animations for games and humanoids

If you have ever built a 3A character controller or tuned humanoid robot gaits, you know the nightmare of animation graphs.

A typical open-world title like Assassin's Creed can easily juggle over 15,000 animations, 5,000 states, and up to 12 levels of nested state machines. One tiny balance tweak can break half your transitions.

NVIDIA wants to burn that old pipeline down.

Researchers from NVIDIA, ETH Zurich, Simon Fraser University, and UT Austin unveiled MotionBricks. Accepted by ACM Transactions on Graphics and heading to SIGGRAPH 2026, it is a real-time motion generation framework that bridges virtual game characters and physical humanoid hardware.

Ditching state machines for smart primitives

MotionBricks runs on a single neural backbone capable of producing over 350,000 distinct motion skills without per-task retraining.

Instead of managing hundreds of transitions, it relies on what the team calls 'smart primitives.' Speed inputs, directional changes, style tags, and in-game collision events are simply turned into sparse keyframe constraints. The neural network fills in the blanks.

Under the hood, it operates as an ultra-fast in-betweening system. A root module first predicts the required duration and trajectory. Then, a pose module samples pose tokens, and a decoder outputs continuous joint motion.

Locomotion supports zero-shot style prompts. You can tell a character to walk like a zombie, limp, or skip, and it blends seamlessly. With its 'Mixture of Styles' capability, players can drift between movements in real time without abrupt animation snap.

For object interactions, like leaping over benches or picking up a sword, the model treats contact points as dynamic keyframes, handling approach and manipulation in one continuous pass.

From Unreal Engine 5 to physical humanoids

Speed is where MotionBricks gets ridiculous.

The framework achieves a throughput of 15,000 FPS and a motion generation latency of just 2 milliseconds between keyframes, beating existing in-betweening baselines across FID, jitter, and reach accuracy.

The team demonstrated high-speed parkour and object manipulation running inside Unreal Engine 5. But they did not stop in the digital world.

MotionBricks was deployed directly onto physical Unitree G1 humanoid robots for full-body locomotion and control. The team explicitly framed this as closing the gap between game animation synthesis and physical robot actuation.

An early preview is already up on GitHub, complete with a lightweight Unitree G1 demo and synthetic training pipeline. A full release, integrated into NVIDIA's GR00T Whole-Body Control framework, is slated for around July 2026.

Why it matters

We think MotionBricks is one of the smartest bridge-building efforts NVIDIA has attempted in the physical AI space.

For game studios, it eliminates the fragile, labor-intensive web of nested animation graphs. Technical animators can stop hand-crafting thousands of state transitions and let a 2ms generative model synthesize natural locomotion on the fly.

For robotics, the implications are even bigger. Real-world physical training data is notoriously scarce and expensive, but game developers sit on decades of high-quality motion capture libraries. By mapping both domains to a shared latent space, robots can essentially inherit virtual athletic versatility for free.

That said, real physics is unforgiving. Synthesizing smooth trajectories at 15,000 FPS in a simulator is not the same as recovering from a slip on uneven pavement. How well this pipeline survives chaotic real-world perturbations will depend heavily on the upcoming GR00T integration in July.