- The reported task count describes current collection progress; the larger hour totals remain two-year targets.
- Simulation evaluation results are company-reported tests, not independently reproduced measurements.
Progress and a larger target
JD has collected data covering more than 100 specific production and service scenarios and over 100,000 real operating tasks, IT Home reported on November 4. The update accompanied a speech by Cao Peng, chairman of JD's technology committee, at the Wuzhen summit's robotics and embodied-intelligence forum.
The company is building embodied-AI data collection centres and plans to gather more than 10 million hours of real-world human video and over 1 million hours of robot operating data within two years. These are prospective collection goals, distinct from the current task count. The report says collection is intended to comply with applicable legal requirements.
Connecting data with testing
JD describes a data-service pipeline spanning collection, storage, annotation, training, evaluation, simulation and testing. Its JoyAI-Sim physical-simulation engine is intended to connect human-operation data, simulated environments and robot evaluation.
In completed company tests, JD reports approximately 90% agreement between simulation and real-world evaluation results, with evaluation efficiency 3.2 times that of physical-robot testing. Those figures are attributed test results; they do not establish equivalent performance in every task or independent validation of the system.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01
