NVIDIA 在 GitHub 发布 MotionBricks,新增 SOMA 人体模型实时动作生成
NVIDIA's MotionBricks Adds a Pretrained Human Model for Real-Time Character Animation From Target Poses
NVIDIA 在官方 GitHub 仓库发布 MotionBricks 实时动作生成模型,在原先仅支持 Unitree G1 机器人的基础上新增 SOMA 人体角色预训练模型,两者共用同一套代码路径。
MotionBricks 在原有 G1 机器人模型之外补上 SOMA 人体模型,并公开权重与运行时代码,读者可据此判断能否接入现有动画管线。
NVIDIA released MotionBricks, the real-time motion generation model from its SIGGRAPH paper, in its own official GitHub repository with a new pretrained model for SOMA human characters, after a first release that covered only the Unitree G1 robot. The model fills in the motion between a character's four most recent poses and up to four target poses, so one runtime can drive keyframing, locomotion and object interactions for a human skeleton or a robot.
- The G1 robot model shipped first. MotionBricks debuted as a Unitree G1 preview inside NVIDIA's GR00T Whole-Body Control project. The new repository adds the SOMA human model and runs both through one code path.
- Your application decides the targets. The same in-betweening model produces keyframe animation, continuous locomotion or object interactions, depending on which target poses it receives.
- NVIDIA claims 15,000 FPS and 2 ms latency. According to NVIDIA's project page, a single backbone covers more than 350,000 motion clips and applies to new tasks without fine-tuning or task-specific tagging.
- The weights permit commercial use. The models fall under the NVIDIA Open Model Agreement, which allows commercial use and redistribution of modified models subject to its terms. The code is Apache 2.0.
- It needs an NVIDIA GPU. The runtime uses TensorRT, and the demos were tested on a GeForce RTX 4090.
Each generation covers up to 1.9 seconds, and the model sets the timing
MotionBricks generates at 30 Hz. The four most recent poses act as context; one target pose sets where the character ends up and stops, while additional targets also set how it moves at that point.
The model decides how long the transition takes, producing 6 to 16 tokens of four poses each, or about 0.6 to 1.9 seconds of motion per generation. Three learned modules split the work:
- Root module. Predicts the duration and the path of the character's root.
- Pose module. A masked token transformer that predicts motion tokens along that path.
- Decoder. A VQ-VAE with several codebook heads that turns the tokens into continuous motion reaching the targets.
Longer motion comes from chaining. The runtime plays the generated motion, records it, and feeds the last four recorded poses back in as context for the next generation.
Both models generate the same way and differ only in pose layout, so one code path drives a human skeleton and a Unitree G1 robot.
NVIDIA says MotionBricks also powers the motion-generation layer of its GR00T Whole-Body Control stack.

Smart Primitives package vaulting, sitting and other interactions as reusable target poses
A Smart Primitive describes an action as target poses placed around an object. You author one in the Smart Primitive Editor, then place it in a scene with the Scene Editor; according to the MotionBricks project page, the backbone fills in the approach, contact and follow-through.
The Scene Preview demo drives a character built on the SOMA body model, or the G1, through a gym scene where it walks, runs, strafes facing the camera, crouch-strafes, vaults, ducks and sits. Six style modes, including Skipping and Zombie, change how it walks.
NVIDIA describes Scene Preview as a distillation of the Unreal Engine demo behind the paper's videos into a small Python runtime. The project page hosts a 2:40 uncut UE5 demo, and NVIDIA says neural networks generated every motion in it, transitions included.
The bundled editors sample target poses from clips, but the repository notes that a motion generation model such as Kimodo or NVIDIA's ARDY could produce those poses instead.

Setup needs a one-time TensorRT build, and sample clips carry evaluation-only terms
The runtime requires an NVIDIA GPU, and the demos were tested on Windows and Linux.
You install Git LFS and uv, clone the repository, and run git lfs pull and uv sync. The first launch of each model and precision builds an NVIDIA TensorRT plan, which takes a few minutes and is cached for later launches.
The Python runtime is built on NVIDIA Warp for GPU kernels, with TensorRT handling inference.
The viewer and scene come from NVIDIA's Newton physics engine.
Four apps ship in the repository: the Inbetweener, which generates motion through poses you place, Scene Preview, the Smart Primitive Editor and the Scene Editor. A guided tutorial walks through all four while building one interaction scene.
The apps run on ONNX models bundled in the repository. You only need the SafeTensors checkpoints from Hugging Face to export your own ONNX models, in FP32 by default or mixed FP16, or to build another runtime.
Both models were trained on Bones RigPlay 1 data, retargeted for the G1. The included sample motion clips carry a separate NVIDIA Sample Data Evaluation License limited to evaluating and testing NVIDIA technologies.
来源:VP Land · vp-land.com