SSyncDev

MotionBricks

SyncDev model directory · Profile updated Nov 19, 2026

Compare with another model

MotionBricks is NVIDIA's open real-time motion generator for SOMA human characters and the Unitree G1 robot, filling motion between context and target poses.

Key specifications

Specification score
4.9
License
NVIDIA Open Model Agreement (weights); Apache-2.0 (code)

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Open weights for self-hosting; no hosted API price. Hardware cost is yours.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input

Produces

  • Text output

Capabilities

Open weights

What is MotionBricks?

Animators call it in-betweening: you set where a character is and where it must end up, and something fills the frames between. MotionBricks does this in real time for two bodies, a SOMA human skeleton and the Unitree G1 humanoid. Feed it the four most recent poses and one to four target poses, and it generates the connecting movement at 30 Hz, deciding for itself how long the transition should take.

The method was published at SIGGRAPH 2026. A G1 preview first appeared inside GR00T Whole-Body Control; the standalone weights repository, with the new SOMA model, went up on 23 October 2026.

Three modules, one generation

  • A root module predicts duration and the path of the root.
  • A pose module, a masked token transformer, fills in motion tokens along that path.
  • A third network completes the pipeline; the three always run together.

A single generation yields 6 to 16 tokens of four poses each, about 0.6 to 1.9 seconds of motion. Chaining generations gives continuous movement. One target pose fixes where the character stops; extra targets also shape how it arrives.

Smart Primitives sit on top. They turn target poses into reusable building blocks for navigation and object interaction, without retraining. The bundled scene demo shows the idea: a character walks, runs, vaults, ducks and sits its way through a level.

What you download

Package Purpose
soma/ Weights and config for SOMA human motion
g1/ Weights and config for Unitree G1
Software repo ONNX models, Warp/Newton/TensorRT runtime, Inbetweener and Scene Preview demos, editors

Keep them apart. Each folder carries a manifest and SHA-256 checksums. These are not Transformers from_pretrained models, and no training pipeline or full training data is included. To just try the demos you can skip the weights entirely, since ONNX builds ship with the software. TensorRT needs an NVIDIA GPU; demos were tested on an RTX 4090 under Windows and Linux.

Robots need more than motion

It outputs kinematics. Not torques. Running it on a real G1 requires a separate tracking controller, robot-limit checks and safety validation. A physics-based controller such as Metatron is the kind of piece that would sit downstream.

Licensing

Weights: NVIDIA Open Model Agreement, which allows commercial use and redistribution of modified models. Code: Apache-2.0. Bundled BONES sample motions carry a separate evaluation-only licence.

To assess it, script ten transitions your game or robot actually needs, such as stop-to-run or crouch-to-stand, and measure foot sliding and target-pose error. Compare with text-driven generators in the text-to-3D directory, and see NVIDIA's wider robotics work on the provider page.

How to evaluate MotionBricks

Compare the model on the work you need to complete.

Tasks and inputs

MotionBricks is listed for Reinforcement learning, Text-to-3D. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under NVIDIA Open Model Agreement (weights); Apache-2.0 (code). Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Real-time motion between the four most recent poses of a character and one to four target poses, at 30 Hz, for SOMA humans or the Unitree G1.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

NVIDIA

Kimodo-SOMA-RP-v1.2

Kimodo-SOMA-RP-v1.2 is NVIDIA's open 282M motion diffusion model that turns text and pose constraints into 3D human animation, now on clean splits.

— context

alexgusevski

alexgusevski/LLaMA-Mesh-mlx

alexgusevski/LLaMA-Mesh-mlx is a source-linked Hugging Face repository assigned to Text-to-3D. Its recorded task interface accepts text and produces 3d.

— context

alexgusevski

alexgusevski/LLaMA-Mesh-q8-mlx

alexgusevski/LLaMA-Mesh-q8-mlx is a source-linked Hugging Face repository assigned to Text-to-3D. Its recorded task interface accepts text and produces 3d.

— context

AvaLovelace

AvaLovelace/BrickGPT

AvaLovelace/BrickGPT is a source-linked Hugging Face repository assigned to Text-to-3D. Its recorded task interface accepts text and produces 3d.

— context