Noitom Robotics, a company specializing in omnimodal data for physical AI, has unveiled HiPHI (High-Precision Human Interaction) at the World Robot Conference in Beijing. This extensive dataset, comprising 617.5 hours of human motion and object interaction from 132 performers, was captured with sub-millimeter precision, making it one of the largest and most accurate publicly available resources of its kind.

The HiPHI dataset is designed to provide a robust foundation for researchers and engineers developing humanoid learning, digital human models, and computer graphics. Its detailed nature supports critical applications such as humanoid tracking, imitation learning, and motion generation, which are essential for creating more natural and responsive AI-driven physical systems. Dr. Lei Han, Chief of Research and Development at Noitom Robotics, emphasized that the dataset's quality and coverage are crucial for humanoids to effectively learn from motion data.

Currently available on Hugging Face under the ModalityNet Open Research License for academic use, with commercial licensing options, HiPHI represents a significant step towards democratizing access to high-fidelity motion data. This move allows a broader range of developers to experiment with and build upon advanced embodied AI concepts, potentially accelerating breakthroughs in areas where physical interaction and nuanced movement are key.

While other datasets exist, HiPHI's scale and precision set it apart, offering a level of detail often reserved for proprietary commercial projects. This public release challenges the traditional model of data exclusivity in advanced robotics research, fostering a more collaborative environment. The ability to train AI policies on such grounded data, as demonstrated by HiPHI-trained policies running on physical Unitree G1 humanoids, highlights its immediate practical utility.