Leju and Zhaofeng Put 15 Humanoids to Work Loading Auto Parts at Hangzhou Plant

For much of the past two years, humanoid robots in Chinese factories have appeared mostly in staged pilots: one machine, a short video clip, a narrow set of parts and a camera crew. The question hanging over the sector has been when, if ever, those demonstrations would settle into ordinary production shifts with real output targets, real quality checks and a payback calculation attached. At an auto-parts plant in Hangzhou, two companies now say that transition has begun.

Zhaofeng Co. Ltd., a Shenzhen-listed auto-parts maker trading under the stock code 300695, and Leju Robotics said on October 15, 2026 that 15 Kuafu humanoid robots have been loading wheel-hub bearings and other components across 15 stations at Zhaofeng's No. 3 factory in Xiaoshan district, Hangzhou, for more than two months. The companies used a joint launch event to present what they describe as an embodied intelligence plus auto-parts loading solution, and reported a 98 percent loading success rate, about 1,000 operating hours and more than 60,000 components handled.

The work is the product of roughly a year of collaboration. The two firms formed a joint venture on October 27, 2025 covering robot-training operations, and the October 15 event marked the point at which the machines moved from pilot testing into routine production support. The components are small: the robots reliably grasp parts with inner diameters of 80 to 120 millimeters, lift them from wire-mesh bins, multi-position stations and loading tables, and place them at the next process step after completing more than 10 motion sequences.

Speed remains the unresolved issue. A loading cycle currently takes under 40 seconds, against about 15 seconds for a skilled worker at the same station, which makes the robot roughly 2.7 times slower per cycle. Zhaofeng's leadership has set a target of 15 seconds within six months to a year, the point at which the company believes the economics would justify going further.

Key Facts

Securities Times reported on October 15 that its reporter visited the No. 3 factory and watched a humanoid bend down, grasp a wheel-hub bearing, turn and place it at the next process step. A Zhaofeng representative told the newspaper that the robots cover 15 factory stations and more than 10 part types, that the loading success rate is stable at 98 percent, and that the machines have run continuously without failure for more than two months, accumulating about 1,000 operating hours while helping produce more than 60,000 components.

Yicai reported on October 16 that the road to that number was long. Zhaofeng chairman Kong Aixiang said he initially assumed a humanoid could be bought, unpacked, charged and used like a phone, but found deployment far harder: at first the robot could not stand, fell over and could not hold parts, and grasping alone took about 15,000 training attempts with repeated adjustments to force and algorithm models. Leju executive vice president Ke Zhendong told the outlet that a success rate above 95 percent is easy in a controlled environment, but real factories add people, noise and changing light.

On the software side, Leju says its development toolchain splits the loading skill into more than 30 basic modules that can be combined without low-level coding, cutting application development from months to weeks. A multi-robot scheduling platform pushes standard skills to multiple machines at once, reducing deployment from days to hours. The line also generates more than 40,000 real-machine operation records per month, and the companies have donated an auto-parts loading dataset to the OpenLET community of the OpenAtom Foundation.

RobotsBeat reported on October 17 that the 98 percent figure is company-reported rather than the result of an independently audited reliability study, and that it is not presented as a per-robot total. The outlet noted that a loading success rate measures something different from uptime or productivity: the disclosures do not say how long a failed attempt takes to recover, how often people intervene, or what share of a shift is spent on useful work.

Robot24 reported on October 17 that flexibility is the commercial rationale. Some Zhaofeng orders are as small as 250 or 300 sets, so switching quickly between components matters more than raw throughput. Zhaofeng general manager Kong Chenhuan said changeovers normally take engineers two to three hours, and noted that recalibrating more than 20 stations outside production hours could add two to three hours of daily output.

Analysis

The headline number here is 98 percent, and it is worth being precise about what it does and does not establish. It says that when a Kuafu humanoid attempts to load a part at one of these 15 stations, it succeeds almost always. It says nothing about the pace of recovery when it fails, nothing about how the line behaves on a bad day, and nothing about whether the same result would survive a move to a different factory with different bins, different lighting and different parts.

The more telling figure is the cycle time. Under 40 seconds against about 15 seconds for a worker means the robots are not yet competitive on speed, and Zhaofeng's own chairman has framed 15 seconds as the gate for expansion, with payback estimated at roughly three to four years only if that target is hit. What this really means is that the deployment is best read as a data-collection exercise with a production function attached, not yet as a labor substitution story. The OpenLET dataset donation and the 40,000 monthly real-machine records point in the same direction: the value being manufactured here is training data as much as it is wheel-hub bearings.

Leju chief executive Chang Lin offered a useful frame for that. He said pretraining models may be 30 to 40 percent accurate, mid- and post-training on real-robot data can reach about 95 percent, and pushing beyond that requires data flowing back from deployments. OpenAtom Foundation deputy secretary-general Li Bo put the scale of the gap in context: training a general-purpose embodied model typically needs tens of millions of hours of real physical interaction data, while only about 500,000 hours of compliant data exist globally.

Seen that way, 15 robots on 15 stations is a small but genuine data pipeline, and the companies' claim that development time has fallen from months to weeks matters more than the success rate. The bigger picture here is that Chinese humanoid vendors are competing less on hardware specifications than on the speed with which they can turn a customer's factory into a training ground and recycle what it learns into reusable skill modules.

Why It Matters

Auto-parts manufacturing is an unforgiving first market. Margins are thin, volumes swing, and orders of 250 to 300 sets force frequent changeovers that punish any automation that takes hours to reconfigure. If humanoids can absorb that variability, they address a real gap that fixed automation has never filled well. If they cannot, they remain a demonstration line with a press release attached.

The public accounting also raises the bar for the rest of the sector. Company-reported success rates, operating hours and component counts are now the standard disclosure package, and the absence of independent verification is a problem the industry will have to solve if buyers are to commit capital at scale. RobotsBeat's point is the relevant one: sustained output, recovery performance and deployment costs across more stations will be the evidence that counts.

There is a competitive dimension too. Leju assistant vice president Ren Guangjie said covering the vast range of manufacturing scenarios will require many ecosystem partners plus more open real-robot data and toolchains, a position that casts the OpenLET contribution as strategy rather than charity.

Next Up

The immediate test is the clock. Zhaofeng wants a 15-second loading cycle within six months to a year, and the current mark of 35 to 40 seconds leaves a wide gap to close. The company is also weighing what a broader rollout would require: more stations, recalibration routines that do not eat into output, and a maintenance and running cost profile that nobody has yet published.

For Leju, the next milestone is whether the modules, the scheduling platform and the data flowing back from Xiaoshan produce measurable gains at a second customer. Until then, the 15 Kuafu machines stand as the most concrete public answer to a question the industry has been asking for two years: what does routine humanoid work in a real factory actually look like, and how fast is it.