For years, the race for 'embodied intelligence'—AI that can perceive and interact with the physical world—was measured in the dexterity of robotic hands or the fluid gait of bipedal walkers. However, a strategic shift is underway among China’s technology giants. The bottleneck has moved from hardware engineering to a desperate hunger for data, marking a transition from a period of technical validation to one of foundational infrastructure building.
JD.com, the Chinese e-commerce titan, recently signaled this pivot by launching JoyEgoCam, a wearable high-definition data collection device designed for the logistics, retail, and healthcare sectors. The goal is to amass 10 million hours of real-world, first-person video data within two years. This aggressive target highlights a stark reality in the industry: while hardware is advancing rapidly, the 'brains' of these robots remain significantly underdeveloped due to a lack of training material.
According to industry leaders like Cao Peng, Chairman of JD’s Technology Committee, training a robot with true generalization capabilities requires at least 10 million hours of data. Currently, the industry is operating on a scale of only hundreds of thousands of hours. This data deficit limits a robot's ability to operate across different environments, keeping embodied AI confined to controlled lab settings rather than allowing it to enter the messy reality of factories and homes.
To bridge this gap, Chinese firms are moving away from traditional 'teleoperation'—where a human manually controls a robot to record movements—because it is too slow and expensive to scale. Instead, they are adopting a 'data pyramid' approach. This model places millions of hours of first-person video at the base, followed by human-robot alignment data in the middle, and a small, high-quality peak of teleoperated precision data at the top.
This shift creates a new competitive logic where the winners are determined by their access to 'edge cases' and diverse commercial scenarios. Companies with massive physical footprints, such as JD.com’s warehouses or Xiaomi’s automated factories, possess a natural advantage. Unlike autonomous driving, which benefits from the 'shadow mode' of millions of cars already on the road, robotics data must be meticulously harvested from real-world business operations.
The competition is also evolving into a battle for industry standards. As firms like Agibot and Xiaomi build their own data service platforms, the race is on to see whose data structure and labeling formats will become the default for the industry. In the global landscape, while Western firms often lead in foundational models and simulation, China is betting that its superior supply chain and vast array of physical data-rich environments will allow it to build a superior 'data flywheel.'