Since the beginning of the year, major companies and startups have significantly increased their investment in data collection. Companies like Qianxun Intelligent, Lingqiao Intelligent, and Lingchu Intelligent have announced ambitious targets for collecting millions of hours of data. Meanwhile, Guanglun Intelligent has completed a financing round of 1 billion yuan, becoming the world's first embodied data unicorn. JD.com has unveiled a comprehensive infrastructure for embodied intelligent data collection, planning to mobilize 600,000 people for crowdsourced data gathering across 64 training sites in 27 cities.
This surge in data collection efforts highlights the industry's focus on building data sets, annotation teams, and simulation environments. However, a critical physical limitation is being overlooked: most teams simplify data collection to perception-level image and point cloud gathering, neglecting the essential motion data from the interaction between robots and physical environments. According to the Guizhou Provincial Big Data Bureau, only 500,000 hours of compliant data from real physical interactions currently exist in China, while the China Electromechanical Integration Technology Application Association estimates that commercializing robotics requires at least tens of millions of hours of data support, indicating a gap exceeding 99% based on a conservative estimate of 10 million hours.
The current challenges in real-world data collection stem from structural constraints that create a physical ceiling. While virtual environments can generate training data at low cost, the gap between simulation and reality is widening as model complexity increases. The AI Index Report 2026 from Stanford HAI reveals that robot manipulation success rates drop from 89.4% in simulated environments to just 12% in real home settings. This discrepancy underscores the need for real physical interaction data, as many robots struggle in unstructured environments like stairs and uneven surfaces, which are crucial for embodied intelligence applications. Continuous data collection is necessary for iterative algorithm development, yet many data collection vehicles are designed for specific scenarios, leading to high costs and inefficiencies in cross-environment deployments.
Editor's Note
The robotics industry is facing significant challenges in data collection, particularly regarding the physical limitations of current technologies. As companies ramp up their data gathering efforts, the disparity between simulated and real-world performance highlights the urgent need for comprehensive physical interaction data. This situation calls for innovative solutions to bridge the gap and enhance the capabilities of robotic systems in diverse environments.
Leave a comment