--- title: "NVIDIA's MotionBricks Replaces Decades of Game Animation Pipelines at 15,000 FPS" subtitle: "NVIDIA's MotionBricks runs 350,000 motion skills from a single neural model at 15,000 FPS, with open-source code and robotics support out of the box" url: "https://alphasignal.ai/news/nvidia-s-motionbricks-replaces-decades-of-game-animation-pipelines-at-15-000-fps" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "NVIDIA AI" company_url: "https://alphasignal.ai/company/nvidiaai" published: "2026-06-11T22:16:38.000Z" updated: "2026-06-12T00:20:00.242Z" type: "News" topics: - "Robotics" - "Open Source" subtopics: - "Vla Models" upvotes: 1165 read_time: "5 min" source_url: "https://www.nvidia.com/en-us/events/siggraph/" comments: 1 --- # NVIDIA's MotionBricks Replaces Decades of Game Animation Pipelines at 15,000 FPS > NVIDIA's MotionBricks runs 350,000 motion skills from a single neural model at 15,000 FPS, with open-source code and robotics support out of the box 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](https://nvlabs.github.io/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](https://huggingface.co/datasets/bones-studio/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 -- `train_vqvae.py`, `train_pose.py`, and `train_root.py` -- letting you swap or retrain parts independently. - **Smart Primitives:** Two high-level interfaces sit on top of the backbone. *Smart Locomotion* takes velocity, heading, and style commands and composes natural movement on the fly. *Smart Objects* let you specify scene interactions as a flexible set of proxy keyframes -- the model fills in the approach, contact, and follow-through automatically, with natural variation across runs. Every motion in the production demo is generated by neural networks -- no foot-locking, no blending, no collision detection, no hand-authored transitions. The 2:40-minute uncut Unreal Engine 5 demo the team published covers the full spectrum of navigation and object-scene interaction needed for industry-level game design. ## Training on human motion, deploying on robots The training corpus is 350,000 production-grade mocap clips from real human actors, retargeted onto the Unitree G1 humanoid robot using a Newton-based solver called the SOMA Retargeter. This retargeting step -- mapping human body proportions onto robot joint constraints -- is what makes the same model usable for both character animation and physical robotics. MotionBricks is now a core organic component of NVIDIA's GR00T Whole-Body Control effort, powering the motion-generation layer of the stack. Together with GEAR-SONIC, it anchors NVIDIA's GR00T Whole-Body Control initiative. That's a meaningful signal: this isn't a standalone research demo, it's load-bearing infrastructure inside NVIDIA's humanoid robotics platform. ## What it's good at -- and where it stops The headline numbers are impressive, but the more interesting claim is qualitative. The team says MotionBricks is the first neural model to reach production-level quality, controllability, complexity, and completeness simultaneously. Side-by-side comparisons against six state-of-the-art in-betweening baselines (models that fill in motion between keyframes) are included on the project page. What it handles well: - Stylized locomotion -- zombie, injured, skipping, strafing -- generated zero-shot from a single prompt - Continuous runtime transitions across speed, direction, and gait, dialed live from user commands - Object interactions: picking up items, jumping over obstacles, sitting, falling - Plug-and-play authoring inside Unreal Engine 5 with no animation graph wiring The current limitations are mostly about scope. The preview release targets the Unitree G1 humanoid specifically. A full release -- a model fully embedded in GR00T Whole-Body Control's robotics formulation, along with the complete training pipeline -- is targeted for approximately one month out. Non-humanoid characters and highly stylized fantasy rigs aren't demonstrated yet. ## Getting your hands on it As an initial preview release, the code is available at the GR00T-WholeBodyControl/motionbricks repo. It ships two components: an interactive G1 demo that lets anyone play with a lightweight MotionBricks-controlled G1 out of the box, and a self-contained synthetic training pipeline with additional instructions for incorporating the BONES-SEED dataset. To get started: ``` git clone https://github.com/NVlabs/GR00T-WholeBodyControl.git cd GR00T-WholeBodyControl git lfs install git lfs pull # fetches ~2.2 GiB of pretrained checkpoints python motionbricks/scripts/interactive_demo_g1.py ``` MotionBricks pretrained checkpoints are skipped by default to avoid an extra ~2.2 GiB download during normal monorepo setup, so you'll need to pull them explicitly. A step-by-step guide to training MotionBricks on your own motion data and adapting it to a new robot is available at `docs/adding_your_own_dataset.md`. The model is free and open-source. ## Why this matters beyond games The industry context here is important. The game animation pipeline has been largely unchanged since the early 2000s -- motion capture, clip libraries, state machines, and hand-tuned transitions. Generative models have been promising to disrupt this for years, but latency and controllability have always been the blockers. A model that runs at 15,000 FPS with 2ms latency removes both excuses at once. The robotics angle is equally significant. SONIC, the humanoid behavior foundation model that underpins GR00T, gives robots a core set of motor skills learned from large-scale human motion data. MotionBricks extends that foundation into real-time interactive control -- meaning the same motion intelligence that animates a game character can now drive a physical humanoid robot responding to live commands. The assumption that needs updating across the field: you no longer need separate pipelines for offline animation authoring and real-time runtime control. A single generative model, trained once on a large enough motion corpus, can serve both roles -- and apparently do it fast enough to run thousands of characters simultaneously. ## Sources - [X post by @nvidiaai](https://x.com/nvidiaai/status/2065196536276173306) - [MotionBricks Project Page -- NVIDIA Research](https://nvlabs.github.io/motionbricks/) - [MotionBricks arXiv Paper (2604.24833)](https://arxiv.org/abs/2604.24833) - [GR00T-WholeBodyControl GitHub Repository](https://github.com/NVlabs/GR00T-WholeBodyControl) - [GR00T-WholeBodyControl Documentation](https://nvlabs.github.io/GR00T-WholeBodyControl/) - [BONES-SEED Dataset on Hugging Face](https://huggingface.co/datasets/bones-studio/seed) ## Comments - **kevin9** (2026-06-12): 15k FPS is wild