NVIDIA expands MotionBricks with human motion models

NVIDIA’s October 27 update adds SOMA human-motion models, downloadable checkpoints and ONNX exports to MotionBricks, alongside an interactive reference implementation built with Newton. Developers guide characters with target poses, while the system generates the movement connecting them. The release contains research code and models; the Unreal Engine integration shown in the research demonstration is excluded.

The repository explains how continuous animation works: the application chooses the targets, and the model fills the gap between four recent poses and up to four target poses. The runtime plays and records the generated movement, then uses its last four recorded poses as the next input context. In the Inbetweener demo, users select poses from clips and place them in the view. TensorRT requires an NVIDIA GPU. The demos use bundled ONNX models, so downloading separate checkpoints is unnecessary for running them.

The SOMA model card sets boundaries for integration. Its inputs represent body poses through joint positions and rotations, together with motion constraints. Text prompts and scene geometry are outside that interface: the application places targets and handles obstacles. Other skeletons need a suitable representation or retargeting step. Targets outside the training distribution can produce lower-quality motion. Results can vary with numerical precision and the inference runtime, with differences accumulating when generated motion becomes later context. Generated movement is not validated for physical robot control.