The project is real and the headline numbers check out against the paper. The abstract states MotionBricks models a dataset of over 350,000 motion clips with a single model and achieves "a real-time throughput of 15,000 FPS with 2ms latency." Section 7.1 describes the primary training set as "a proprietary motion capture collection containing approximately 700 hours of high-quality motion data with 350k motion clips," covering "9,300 unique skills across 36 categories, captured from over 163 performers." Section 7.4 reports the timing figure "on a desktop with an RTX 5090 GPU," and the contributions list states the authors "deploy MotionBricks on the Unitree G1 humanoid robot."

The release picture is narrower than the post describes. The NVIDIA project page describes the GitHub code as "an initial preview release" shipping two components: an interactive demo with "a lightweight MotionBricks-controlled G1 out of the box," and a self-contained synthetic training pipeline with instructions for incorporating the BONES-SEED dataset. The same page and the repo README state that a full release, meaning a model fully embedded in GR00T Whole-Body Control's robotics formulation plus the complete training pipeline, is "targeted for approximately one month out," with reproducibility experiments in flight. The repo changelog dates the preview to 2026-04-27. As of 2026-09-01, roughly four months later, the repo still carries that pending-full-release status.

The dataset is owned by Bones Studio, not NVIDIA. The publicly available BONES-SEED release is described on Hugging Face as 142,220 annotated human motion animations, and NVIDIA's own GR00T documentation changelog logs it as "142K+ human motions (~288 hours)." Access is gated: the Hugging Face card requires reviewing conditions before access and directs commercial parties to licensing@bones.studio, and the Bones Research Network page describes the free click-to-agree tier as available to non-profit institutions only. The license defines the data as proprietary and confidential and bars using it to train generative models whose output functions as a commercial substitute for motion-capture data, while expressly permitting training control policies and similar models.

There is an internal inconsistency in the public record worth flagging: NVIDIA's project page and repo describe BONES-SEED as "MotionBricks' training corpus, 350k production-grade mocap clips," while the paper calls its 350k primary set proprietary and the dataset owner publishes 142,220 sequences at roughly 288 hours. These are not the same quantity, and I could not resolve which artifact corresponds to the 700-hour figure.

On the "one AI for both" point, the paper says of the game and robot demonstrations: "For both the UE5 and G1 applications, we use the same model architecture and training settings." That is shared architecture and recipe. The paper also lists the G1 skeleton as a distinct 34-joint configuration, uses a separate LaFAN1-G1 retargeted benchmark, and NVIDIA's repo describes a "SOMA Retargeter" that "retargets SOMA capture onto the G1, producing MotionBricks' training data," while the UE5 demo retargets output onto game characters at runtime. Nothing I retrieved states that a single set of weights drove both the game character and the robot.