The real challenge in building robots that can move like humans isn't computing power or clever algorithms,it's access to high-quality motion data that robots can actually learn from. Noitom Robotics just released HiPHI, a publicly available dataset containing 617.5 hours of precisely captured human movement and object interactions, marking a significant shift in how the physical AI (embodied AI) community approaches robot training.

Why Is Motion Data Such a Bottleneck for Robot Learning?

The physical AI field has long faced a frustrating trade-off. Internet-scale video datasets are abundant but lack the precise physical measurements robots need to actually learn from them. Meanwhile, laboratory-captured motion data is incredibly accurate but covers only narrow slices of human behavior and is typically kept as proprietary company assets. High-precision motion capture is slow and expensive to produce, which is why public releases at scale are exceptionally rare.

HiPHI breaks this pattern by making 617.5 hours of motion data freely available to researchers. The dataset includes 371.8 hours of whole-body human motion and 245.7 hours of human-object interactions, all captured from 132 performers at 90 Hz with sub-millimeter precision. Each object's trajectory and mesh are recorded in sync with the performer, creating a foundation that robots can actually learn from.

"The bottleneck in physical AI is not how much data exists, but how much of it a machine can actually learn from. HiPHI is our first answer: coverage designed before a single frame was captured, precision that survives the transfer to real robots, and the objects people touch recorded as part of the motion itself," said Dr. Tristan Ruoli Dai, Founder and CEO of Noitom Robotics.

Dr. Tristan Ruoli Dai, Founder and CEO of Noitom Robotics

The dataset was released at the World Robot Conference in Beijing and is now available on Hugging Face, a platform where researchers share machine learning models and datasets. Commercial licensing is also available through ModalityNet.

How Does HiPHI Actually Help Robots Learn Better?

The dataset is organized around motion units drawn from FrameNet, a linguistic framework that organizes human action semantics. This structure matters because it helps robots understand not just how humans move, but the semantic meaning behind those movements. Policies trained on HiPHI have already been deployed on physical Unitree G1 humanoid robots, demonstrating the dataset's real-world applicability. These robots successfully learned to run, sit, crawl, carry a box, and pull a suitcase.

What makes HiPHI particularly valuable is its quality metrics. The research team published comprehensive documentation including an academic paper that reports the broadest measured motion coverage and the lowest per-frame physical-quality error values among comparable datasets. This transparency is unusual in a field where data quality is typically a closely guarded competitive advantage.

• Coverage and Breadth: 617.5 hours of motion data organized around semantic action units, providing diverse human behaviors for robot learning
• Precision and Accuracy: Sub-millimeter marker tracking at 90 Hz, with synchronized object trajectories and mesh data that survives transfer to real robots
• Transparency and Documentation: Full academic paper, published quality metrics, and evaluation guides that allow researchers to verify the dataset's reliability

What Does This Mean for the Physical AI Industry?

The release signals a broader shift in how the embodied AI community approaches development. Rather than treating motion data as a proprietary moat, Noitom is positioning high-quality datasets as a shared foundation that benefits the entire field. This approach mirrors how large language models (LLMs), which are AI systems trained on vast amounts of text, have accelerated progress in natural language processing by enabling researchers to build on common baselines.

The timing is significant. Across the industry, companies are racing to deploy humanoid robots in real-world scenarios. Faraday Future recently announced that its subsidiary AIxC is pivoting entirely to physical AI and robot-sharing operations, with RoboShare completing its first paid commercial order involving multiple robot form factors including humanoid and quadruped robots. Meanwhile, researchers at institutions like Tsinghua University and Shanghai Jiao Tong University are developing foundational vision-language-action models and multi-agent systems that depend on high-quality training data.

"Coverage, per-frame quality, and object state decide whether a humanoid can actually learn from motion data, so those are the things we designed for, measured, and published. Everything a researcher needs is in the release: standardized BVH, synchronized object trajectories, a semantic motion index, and the evaluation guide," explained Dr. Lei Han, Chief of Research and Development at Noitom Robotics.

Dr. Lei Han, Chief of Research and Development at Noitom Robotics

Noitom's infrastructure produces more than 100,000 hours of motion data annually for commercial partners, suggesting that HiPHI is just the beginning of a larger effort to make high-quality motion data more accessible. The company plans further public releases, beginning with an omni-modality interaction corpus later this year, and upcoming releases will add support for widely used formats such as SMPL and SOMA.

For the physical AI field, this represents a critical inflection point. As robots move from laboratory demonstrations into commercial operations, the quality of their training data will increasingly determine whether they can handle real-world complexity. By making HiPHI public, Noitom is essentially raising the baseline for what counts as acceptable motion data in the industry, potentially accelerating progress across the entire ecosystem.