15 Humanoid Robots Log 1,000 Hours on Auto Parts Line: 98% Success Rate and 40,000 Data Points a Month

Forum topic · QianXun · 2026-10-15

Summary

Fifteen humanoid robots from Shenzhen-based Leju Robotics are now working full-time on an automotive parts loading line at Zhejiang Zhaofeng, a Chinese auto components manufacturer. The system handles preparation, picking, and transfer to 15 workstations across more than 10 part types, with each cycle kept under 40 seconds. The companies report a stable 98% task success rate, roughly 1,000 cumulative hours of operation, and over 60,000 parts produced. The architecture rests on three tools: a skill toolchain that breaks loading into 30+ reusable modules (cutting development from months to weeks), a multi-robot scheduling platform that syncs skills to all units in hours instead of days, and a data loop where every shift collects training data without stopping production — over 40,000 entries per month. That real-machine dataset has been donated to the OpenLET community under the OpenAtom Foundation. The project, announced October 15, is positioned as China's first replicable embodied-AI field training demonstration, aligned with a June MIIT initiative targeting normalized deployment by end of 2026.

🏭 New Colleagues on the Line

The automotive parts production line at Zhejiang Zhaofeng Co. recently gained fifteen tireless coworkers: humanoid robots from Shenzhen's Leju Robotics. Their job is material loading — from preparation and picking to transfer and delivery of parts to workstations. Each cycle stays under 40 seconds, and the system adapts to 15 workstations and more than 10 part types.

On October 15, the two companies jointly released the "Embodied AI + Automotive Parts Loading Solution" in Zhejiang. Industry observers have framed it as the "industry's first replicable embodied intelligence real-world training demonstration project" — in other words, humanoid robots moving from stage demos to actually working on the line.

Two numbers outweigh any adjective in the press release: a task success rate stable at 98%, and roughly 1,000 cumulative hours of operation assisting the production of over 60,000 parts. Demo videos can be edited; 1,000 hours cannot.

🔄 A System That Grows Itself

The real engineering effort here turns "using robots" into "cultivating data."

Modular skills: The skill development toolchain decomposes loading actions into 30+ basic modules that combine like building blocks — no low-level code needed, cutting development cycles from months to weeks.
Multi-robot scheduling: A dispatch platform syncs standard skills to all units with one click, shrinking deployment from days to hours.
Safety modules: Part-drop, collision, and overspeed detection enable reliable continuous operation.

The data forms a closed loop: collection happens during normal operation (no downtime or dedicated production runs), data is quality-checked and classified, then fed back into model iteration — including real samples from anomalous conditions, which push the success rate higher. Fifteen robots working in parallel collect over 40,000 data entries per month while they work.

This real-machine dataset has been donated to the OpenLET community under the OpenAtom Foundation, so the whole industry can use it to train models. A company donating its production line data amounts to open-sourcing the most expensive thing there is: the field site itself.

🧱 From One Unit to One Fleet

Context: in June, China's Ministry of Industry and Information Technology (MIIT) and the State-owned Assets Supervision and Administration Commission (SASAC) launched a special action program for humanoid robot and embodied AI real-world training, requiring application validation and normalized deployment in representative scenarios by the end of 2026. The Leju–Zhaofeng project sits squarely on that policy line.

Ren Guangjie, assistant VP at Leju, was blunt: the solution can already replicate across workstations and scenarios, but covering the vast landscape of manufacturing scenarios still requires many ecosystem partners across the supply chain to fill in the full deployment puzzle. One line running well does not equal ten thousand lines running well — an entire ecosystem sits in between.

For engineers, the reference value is structural: skill modularization solves slow development, multi-robot scheduling solves slow deployment, and data feedback solves "degrading with use." Three pain points, three tools — homework worth copying.

| Metric | Value | Notes | |---|---|---| | Robots | 15 | Normalized operation on one line | | Cycle time | Under 40 s | Preparation through loading | | Success rate | 98% | Task success, company-reported | | Cumulative runtime | ~1,000 h | Per-unit or total not specified | | Output | 60,000+ parts | Assisted production | | Monthly data | 40,000+ entries | 15 robots in parallel | | Skill modules | 30+ | Building-block composition | | Deployment time | Days → hours | Multi-robot scheduling platform |

Sources and Caveats

Event and figures: consistent across China Securities Journal, Tencent News, Sina Finance, and North News (Oct 15); originally from the joint Leju–Zhaofeng announcement.
"Industry's first replicable" is the press release's own framing, quoted as such; "first" was not independently verified.
The statistical basis for the 98% success rate (retries, denominator) and whether the 1,000 hours are cumulative or per-unit were not detailed in the source materials.
Policy references come from press summaries of the June MIIT/SASAC action program; the original policy text was not reviewed.
Robot model and navigation/manipulation technical details were not disclosed and are not speculated on here.