Leju Intelligent (Shenzhen) Co., Ltd. ("Leju Robotics") and Zhaofeng recently unveiled a joint "embodied AI + auto parts loading solution" in Zhejiang, putting 15 humanoid robots into routine operation on Zhaofeng's auto parts production line. Industry observers call it the first replicable embodied AI real-world training demonstration project in the sector, marking a shift for humanoid robots from "pilot validation" to "everyday production work." In June this year, the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission launched a dedicated campaign on humanoid robots and embodied AI real-world training, setting a target of completing application validation and routine deployment in representative scenarios by the end of 2026.

Covering the Full Prepare–Pick–Transfer–Load Workflow

On Zhaofeng's line, the 15 humanoid robots execute more than 10 action sequences in tight, complex surroundings, reliably gripping small components and handling the entire "prepare–pick–transfer–load" workflow. Each cycle takes under 40 seconds, the system covers 15 workstations and more than 10 types of parts, and the success rate holds steady at 98%. The robots have logged roughly 1,000 hours of operation and helped produce more than 60,000 components.

Behind that sits a system spanning development, deployment, operation, data and safety, according to the companies. A skills development toolchain uses "building block" style composition to cut development cycles from months to weeks. A multi-robot scheduling platform allows one-click skill synchronization, shortening deployment to a matter of hours. A one-click material call system can start multiple machines at once. Operational data from the floor is fed back throughout the process, paired with safety modules that detect dropped parts, collisions and excessive speed to keep the robots running reliably.

More Than 40,000 Operational Data Records a Month Feed Model Iteration

With 15 robots working in parallel, the project has built large-scale concurrent data collection, said an executive responsible for the deployment: more than 40,000 data records a month, accumulating continuously, so the robots get smarter the more they are used.

The data feedback loop has three stages. First, collection happens during work itself, with no need to halt production or schedule separate runs; data is gathered, returned and archived in sync. Second, the data is quality-checked and classified, then routed by intended use. Third, it feeds back into model iteration, steadily lifting success rates while capturing real-world samples such as abnormal operating conditions. The result is a virtuous cycle running from on-site data return to quality checks and classification, model iteration and higher success rates. The dataset of real-robot operational data has been donated to the OpenAtom Foundation's OpenLET community, lowering R&D barriers across the industry.

Ecosystem Collaboration Remains the Key to Scaling

Industry observers say the technology package can already be replicated across workstations and scenarios, but deploying humanoid robots at scale will depend on coordinated effort across the industrial ecosystem.

Ren Guangjie, assistant vice president at Leju Robotics, said that truly covering the vast range of manufacturing scenarios will require large numbers of ecosystem partners from across the supply chain to join in and fill out the capability gaps at every stage of deployment. Leju Robotics said its next step is to continue open-sourcing resources including real-robot data and development toolchains, and to build out the OpenLET embodied AI open-source dataset community to support developers and startup teams. The aim, the company said, is to "validate one, deploy a batch, and lift a whole field," bringing more high-value industrial scenarios into production. Moving from single-unit validation to batch replication, the project is mapping a workable path for scaling humanoid robots in industrial settings and supporting the development of new quality productive forces and higher-quality manufacturing.