Shanghai's Maniformer Surpasses 1 Million Hours of Physical AI Data to Tackle Robotics Bottleneck

Maniformer, a startup in Shanghai that focuses on AI data, has made significant progress in embodied AI development by delivering more than 20,000 of its MEgo-series data collection devices and expanding its dataset to over one million hours of real-world human physical data.

The industry's main bottleneck is high-quality data availability as humanoid robotics and physical AI grow in China's tech sector.

Unitree Robotics founder Wang Xingxing stressed at the recent World Robot Conference that data drives embodied intelligence, but high-quality human demonstration data and full-body motion metrics are scarce.

"Crossing the 20,000-unit threshold marks the shift of embodiment-free data collection into full-scale industrial deployment," said Yao Maoqing, chairman and chief executive of Maniformer and Partner at AgiBot.

"Reaching one million hours is just the foundation – we expect that figure to scale rapidly to 10 million and ultimately 100 million hours as physical AI expands across real-world applications."

Maniformer solves this problem by separating data collection from robot hardware. The wearable MEgo-series devices allow humans to capture complex behavior and manipulate objects in homes, factories, retail stores, and warehouses without a robot.

Maniformer established a base in Shanghai's Zhangjiang Robot Valley to speed processing. Gripping, sorting, and assembly tasks generate rich multimodal data streams that the facility processes using vision, depth, tactile feedback, and spatial trajectories.

The raw inputs are converted into training-ready datasets, reducing labor and hardware costs for robot training data collection.