NVIDIA's MotionBricks now runnable: 1.3–2.5ms per call, but it's a motion generator, not a controller

maier_ak · x · 2026-11-20

Andreas Maier reviews NVIDIA's open-source release of MotionBricks (SIGGRAPH 2026, with SFU, UT Austin, ETH Zurich): a real-time inbetweening model that fills in motion between poses for both game characters and the Unitree G1 humanoid.

How it works: takes the four most recent poses plus 1–4 targets, generates connecting motion at 30fps; one call yields 0.6–1.9 s with model-decided duration, longer motion via chaining. Three learned modules split the work: a root module predicts hip path and duration, a masked transformer predicts discrete motion tokens.

Performance: 1.3–2.5 ms per request on an RTX 4090. Code is Apache 2.0, models commercially usable under NVIDIA's open model agreement; pretrained human and G1 models, a Python runtime, demos and editors are out.

Limits: no text or scene input (the app places targets and avoids obstacles), output is kinematic and unvalidated for real robot control, and no quality benchmark of the released models yet. Verdict: a fast motion generator, not a controller.

Related event: NVIDIA Open-Sources MotionBricks, a Real-Time Motion Inbetweening Model(6 posts)→

Original post →

More from Embodied

Embodied channel →