At the 2026 World Robot Conference Forum, held in Beijing from August 19-22, JD.com's Technical Committee Chairman and JD Cloud President, Cao Peng, delivered a keynote address on the current state and future trajectory of embodied intelligence.
Cao Peng stated that robotics remains on the cusp of a breakthrough, primarily because robots and embodied models still lack sufficient capabilities in task comprehension and generalization. He noted that in certain pilot environments, including select operations within JD Logistics, robots have achieved performance parity with human workers on standardized tasks, with efficiency and cost metrics closely approaching those of manual labor.
However, he emphasized that these successes cannot be scaled to broader, more complex environments. The fundamental constraint lies in poor model generalization, which stems from insufficient training data. As a result, robots are currently confined to specific, narrowly defined use cases and cannot achieve scalable replication across diverse settings.
Industry consensus holds that training a robot "brain" with strong generalization abilities capable of large-scale deployment requires data on the order of ten million hours. Yet, the industry currently supplies only roughly 100,000 hours of training data, creating a massive data deficit. Additional challenges include difficulties in transferring data collected from one robotic platform to another, complicating cross-embodiment adaptation.
Cao Peng argued that the future of robotics hinges on the brain's generalization capacity, which in turn depends on the availability of abundant, high-quality data for model training. To address this, JD.com has committed to launching a global-scale data collection initiative. Over the next two years, the company aims to gather more than ten million hours of high-quality behavioral data from real-world human scenarios, alongside over one million hours of operational data from robotic systems.
This initiative will engage a diverse range of participants, including JD employees across various roles, external industry professionals, and urban residents, with plans to mobilize more than 100,000 individuals. All data collection will strictly adhere to legal and regulatory frameworks, ensuring data security and reliability throughout the process.
Cao Peng expressed hope that this program will establish a sustainable data infrastructure for the embodied intelligence sector, providing a foundational resource that could accelerate advancements across the industry.
Comments