NVlabs/motionbricks

MotionBricks is the official implementation of the SIGGRAPH 2026 paper: the model takes the character's most recent 4 pose frames and 1–4 target poses, generating motion between them via three modules—root trajectory, pose, and decoding—producing approximately 0.6–1.9 seconds per segment, which are concatenated to form long actions. The repository provides pretrained models for the SOMA human body model and the Unitree G1 (ONNX files are included in the repository, while original weights are hosted on Hugging Face), along with a Python runtime based on Warp, Newton, and TensorRT, two demos for pose interpolation and scene preview, and two editors.

MotionBricks is the official implementation of the SIGGRAPH 2026 paper: the model takes the character's most recent 4 pose frames and 1–4 target poses, generating motion between them via three modules—root trajectory, pose, and decoding—producing approximately 0.6–1.9 seconds per segment, which are concatenated to form long actions. The repository provides pretrained models for the SOMA human body model and the Unitree G1 (ONNX files are included in the repository, while original weights are hosted on Hugging Face), along with a Python runtime based on Warp, Newton, and TensorRT, two demos for pose interpolation and scene preview, and two editors.

Why it matters By treating target poses like climbing over or sitting down as "intelligent primitives" that can be assembled like building blocks, there is no need for retraining; the same runtime drives both the SOMA human body model and the Unitree G1 robot.

Star count 1075 (+16 / 24h)

Reasons

  • 16 new stars in the last 24 hours, total stars 1075 (indicating attention only); code licensed under Apache-2.0
  • The README includes links to the project page, paper, and step-by-step tutorials; the demos show characters walking, running, climbing over, crouching, and sitting in scenes

Usage

  • After installing Git LFS and uv, clone the repository, run git lfs pull and uv sync, then run uv run python -m app.scene_preview --scene assets/scenes/g1_gym.scene.json to open the G1 scene preview

Risks

  • The runtime depends on TensorRT, requiring an NVIDIA GPU; the author tested it on an RTX 4090; building TensorRT plans for each model and precision level takes several minutes on first launch
  • Model weights are licensed under the NVIDIA Open Model Agreement, which differs from the Apache-2.0 license for the code; review the terms before commercial use

Open

Trendshift(GitHub Trending 快照)

Explore the latest AI Radar updates