Noitom Open-Sources HiPHI Motion Capture Dataset Alongside NVIDIA Isaac Video-to-Data and SONIC Full-Body Control Model

During the 2026 World Robot Conference (WRC), the embodied AI field received three interlocking open-source releases on the same day: motion capture data (Noitom HiPHI), a video-to-training-data toolchain (NVIDIA Isaac Video-to-Data), and a full-body control model (NVIDIA SONIC). Together, they form the first industrial-grade open reference implementation of the closed loop "where data comes from → how data is used → how models are trained."

1. Noitom HiPHI: 617.5 Hours of Sub-Millimeter Optical Motion Capture

Noitom Robotics officially released and open-sourced the HiPHI (High Precision Human Interaction) dataset during WRC, with the data card live on Hugging Face. Three key numbers:

• 617.5 total hours, 200.1 million frames
• 132 motion capture actors, 90 Hz capture frequency, sub-millimeter optical precision
• Split into 371.8 hours of whole-body human motion + 245.7 hours of human-object interaction (HOI) data (including left-right mirroring)

Three fundamental differences from prior human motion datasets:

1. Precision tier. This is professionally captured data, not approximate poses extracted from internet video via pose estimation. Sub-millimeter precision means every joint rotation and finger movement can be mapped back into robot joint space. 2. Synchronized "motion + interaction." The HOI portion provides synchronized object trajectories and corresponding OBJ meshes. When a person picks up a cup, sits down, or pushes a door, the robot sees not just the motion but the exact 3D position change of the object—foundational for training robots to understand physical cause and effect. 3. Data organization. Structured by FrameNet action semantics into "action units," directly targeting VLA (vision-language-action) model training needs.

Commercial use requires separate licensing (the ModalityNet Open Research License covers research only), but for academia and the open-source community, this is currently the highest-precision, largest-scale, best-organized open human motion dataset available. Notably, models trained on HiPHI have been deployed on the real Unitree G1, performing running, sitting, crawling, box carrying, and luggage pulling—the sim-to-real gap was significantly compressed.

2. NVIDIA Isaac Video-to-Data: Turning YouTube Videos into Robot Training Data

Where Noitom takes the "hand-crafted precision" route, NVIDIA takes the "bulk refinement" route. The same week, NVIDIA open-sourced the Isaac Video-to-Data toolchain, complementary to HiPHI—it automatically converts real human demonstration videos into simulation assets and training data, deployable back onto physical robots, closing the "real → sim → real" loop.

The toolchain includes four core modules:

• Vision Ingestion Agent: automatically segments long videos and retrieves action clips
• Hand/object/body reconstructors: recover hand motions, 6-DoF trajectories, and depth meshes
• Simulation retargeting: retargets human motion onto arbitrary robot bodies and trains RL policies in Isaac Lab
• CHORD (Contact Wrench Guidance From Human Demonstration): uses contact forces to guide dexterous manipulation

Open-sourced alongside: a motion library of 1,215 tasks, 121,500 robot episodes, and the CHORD training framework. For small and mid-sized teams without mocap equipment or real-robot deployment, this is the first toolchain that turns "video" directly into a "training set."

3. SONIC: One Model Controls the Whole Body, 42M Parameters, 700 Hours of Mocap

NVIDIA simultaneously released SONIC (Simulating Open-source Neural Intelligence for Control), an open-source full-body control model. Three numbers define its capability:

• Training data: 700 hours of mocap, 100M+ motion frames
• Model size: up to 42M parameters
• Compute: 128 GPUs, ~21,000 GPU-hours

SONIC's core innovation is a "universal token space"—different modalities (VR teleoperation, video, text, music) are encoded into shared discrete tokens fed into a unified robot control decoder. One model handles "dancing to a video," "generating motion from music," and "following text instructions."

Real-robot results: 99.2% success rate across 123 real motion sequences on the Unitree G1; when connected to a VLA model, 75% average success on five whole-body tasks (object retrieval + hand-foot coordination). This is industrial-grade sim-to-real transfer.

4. What the Three Together Mean

For the past three years, humanoid robotics has been trapped in a chicken-and-egg problem: policies trained in simulators don't work on real hardware, and working real-robot policies require a separate controller per skill. HiPHI + Video-to-Data + SONIC answer three questions simultaneously:

• Where does high-quality real-captured data come from? (HiPHI)
• How can video data be converted to training data cheaply? (Video-to-Data)
• Can full-body control avoid one controller per skill? (SONIC)

The implication for China's embodied AI sector is also clear: during WRC, Noitom open-sourced core data assets (with commercial licensing reserved), while NVIDIA fully opened its toolchain and model—the full stack of "data + toolchain + control model" is being rapidly laid down by leading players under a dual "open-source + commercial license" structure. Chinese embodied AI companies that focus only on robot hardware risk being stuck in the position of "other people's training data + other people's control models + our own hardware manufacturing."

Sources: IT之家 2026-08-22 (Noitom HiPHI release), AIQUEST coverage, NVIDIA official GitHub + Hugging Face, Nivlecreto technical analysis, Guangming Online 2026-08-24 (Beijing Zhongguancun robot training base).