Experts Talk Robotics
On the afternoon of August 12, 2026, Shenzhen Zibianliang Robot Company hosted a live stream.
A robot equipped with dual robotic arms and standard grippers began continuous sorting in the face of chaotically stacked packages. Cardboard boxes, express soft bags, cylindrical parcels, and foam-packaged items varied entirely in size, weight, and material, with completely random placement postures. The robot needed to pick up the packages, determine the position of the shipping labels, flip the labels to face upwards, and then place them on the conveyor belt—all without any human intervention.
An hour later, the results were finalized: 1,816 items, with an accuracy rate of 98%.
The previously publicly demonstrated record for the same scenario by the US-based Figure AI was 1,248 items/hour; Zibianliang exceeded this by 45%. Moreover, according to Zibianliang's internal estimates, the hardware cost of its entire logistics sorting system is only about 30% of Figure AI's solution.
Even more intriguing is the difference in hardware approaches. Figure AI uses a complex solution of "humanoid robot + five-finger dexterous hand," whereas Zibianliang uses only "dual robotic arms + standard grippers"—no dexterous hands, no "humanoid" structure, resulting in a direct 70% reduction in hardware costs.
This live stream sent a clear signal: what determines whether a robot can do the job is not how dexterous its hands are, but how capable its "brain" is. The "brain" behind Zibianliang is WALL-B, the world's first embodied AI large model based on the "Unified World Model" architecture, released in April this year. It integrates vision, language, touch, action, and physical prediction into a single neural network, compressing the "perception-decision-execution" closed loop for a single package to tens of milliseconds.
However realistic Zibianliang's live stream was, it was merely a test. The real question worth asking is: which humanoid robots have already entered real-world scenarios, working day in and day out?
A List of Those "Actually Working"
The following list is compiled from public reports, with only one screening criterion: continuous deployment in real production/operational scenarios, not exhibition demonstrations, not short-term training, and not staged photos.
Who Is Working? Seven Brands, Two Approaches
A total of seven brands appear on the list: UBTECH, Galbot, Agibot, Robot Era, Zoomlion, Leju Robotics, and Wonen Embodied.
These seven companies can be roughly divided into two categories.
The first category is "body manufacturers developing their own brains"—Agibot developed the Qiyuan large model, Robot Era developed ERA-42, Zoomlion developed the Yungu large model, and Galbot developed Grocery VLA. Their commonality is that they build both the body and nurture the brain, pursuing hardware-software integration.
The second category is "body + external brain"—UBTECH integrates DeepSeek-R1 from DeepSeek (with its self-developed cerebellum responsible for motion control), Leju Robotics partners with Huawei's Pangu large model, and Wonen Embodied relies on technology from the Institute for Interdisciplinary Information Sciences at Tsinghua University. These companies choose to entrust large model capabilities to AI partners who excel in this area, while focusing on body engineering and scenario implementation.
Neither approach has absolute advantages or disadvantages, but one trend is clear: self-developed embodied large models are becoming the mainstream. The reason is easy to understand—the data closed loop in factory scenarios is too critical. The tactile and visual data generated by robots working every day must be able to feed back into their own models, forming a flywheel of "getting smarter the more they work."
Where Are They Working? Four Major Scenarios Emerge
Reviewing 15 deployment records, the implemented scenarios are concentrated in four directions:
Automotive factories are the most popular testing grounds. UBTECH has entered Zeekr, BYD, and Foxconn; Leju has entered FAW Hongqi, BAIC, and AITO; additionally, Xiaomi is in XPeng factories, Kepler is in SAIC-GM, and Fourier is in SAIC-GM—automotive production lines have almost become a "required course" for humanoid robots. The reason is simple: workstations in automotive factories are relatively standardized, and processes like handling, quality inspection, and assembly are highly repetitive. Moreover, automakers inherently possess an automation DNA and have high acceptance rates.
Logistics sorting is the scenario with the fastest efficiency gains. Robot Era achieved a peak of 1,200 items/hour in over a dozen SF Express/Postal logistics centers, and Wonen Embodied reached a peak of 1,200 items/hour and a stable 800 items/hour at the Guangzhou Postal District Center—these two figures are already close to the 900 items/hour level of stable manual sorting. Coupled with the 1,816 items/hour measured in Zibianliang's live stream, logistics sorting is becoming the first track where humanoid robots have truly achieved PMF (Product-Market Fit).
Unmanned retail is the only scenario achieving 24/7 uninterrupted operation. Galbot's unmanned pharmacy in Beijing has been operating for over a year, with the robot on duty for more than 150 days, completing tens of thousands of orders with a 95% grasping success rate. In a 50-square-meter space with 5,000 types of goods and over 10,000 boxes of medicine, the robot handles all autonomous sorting and packaging. The special value of this scenario lies in its enclosed space, fixed categories, and 24-hour demand—it is naturally suited for robots.
Deploying in their own heavy equipment factories is a "keep the good stuff in the family" approach. Zoomlion has moved 8 models of humanoid robots into over 300 production lines in its own smart industrial city, with dozens working simultaneously across multiple stations in machining, logistics, assembly, and quality inspection. Validating in their own factories first before pushing to the market—this is the traditional move for equipment manufacturing giants, and it also indicates that humanoid robots have entered the "real combat" stage.
Working Efficiency: The Climb from 20% to 90%
The efficiency data in the list draws a clear climbing curve:
In less than two years, UBTECH climbed from "only 20% of human efficiency" to "doubling efficiency"; Leju improved from 30% to 50%; Robot Era approached 90% of human efficiency in logistics scenarios—the slope of this curve is steeper than many expected.
However, we must remain clear-headed: apart from logistics sorting and unmanned retail, robot efficiency in most factory scenarios still hovers in the 50-70% range of human efficiency. This means a single robot cannot yet independently replace a worker, and human-robot collaboration is often required to make it work. Agibot's Genie G2, where "two units working in tandem can completely replace one worker," is both a milestone and an indication of the current ceiling for single-machine capabilities.
End Effectors: Dexterous Hands Are Mainstream, but Grippers Are Making a Comeback
Among the 15 records, 13 used dexterous hands, with only 2 from Galbot using grippers/suction cups.
The advantages of dexterous hands are obvious: high degrees of freedom, fine tactile perception, and the ability to perform complex operations like screwing, threading needles, and fine assembly. In factory scenarios, dexterous hands are almost standard equipment.
However, Zibianliang's live stream on August 12 poured cold water on the dexterous hand route: using simpler grippers and relying on smarter models actually achieved higher efficiency. Zibianliang partner Gan Ruyi made a counter-intuitive statement: "Using simpler robotic arms + grippers for logistics sorting places higher demands on the large model." In other words, it's not that dexterous hands are bad, but if you have a sufficiently strong brain, you don't need complex hands to compensate for the brain's shortcomings.
If this logic holds, the high cost, high maintenance, and high failure rate of dexterous hands will become a burden for industrial deployment. At least in specific scenarios like logistics sorting, "simple hardware + strong model" may be a more cost-effective route.
Unitree, the Most Popular Brand, Why Doesn't It Have a Single Robot on the Factory Floor?
There is an obvious absentee on this list: Unitree Robotics.
As the "first humanoid robot stock," valued at 60.99 billion RMB, and the company ranking first globally with 5,511 humanoid robot shipments in 2025, Unitree did not secure a single spot on the regular deployment list. Its robots have been tested in Geely factories and carried bricks on Mount Tai, but all remained at the pilot stage, without entering the daily production lines of any factory.
This is not an oversight, but a structural issue.
First, 73.6% of revenue comes from scientific research and education, with industrial scenarios accounting for only 9%. Unitree's customers are university laboratories, not factory workshops. Scientific research customers buy prototypes for algorithm research and do not have high requirements for continuous operation stability; factories need equipment that can work non-stop for 8 hours under real working conditions. The technical requirements for the two are not on the same order of magnitude. Data shows that the paid conversion rate for Unitree's pilot projects with Great Wall Motor and Tokyo Haneda Airport is less than 10%.
Second, a well-developed cerebellum but an underfunded cerebrum. Unitree's technical accumulation focuses on motion control—backflips, running, and dancing. These "cerebellum" capabilities are indeed world-class. But what factories need is "cerebrum" capability: environmental perception, task planning, and scenario generalization. Unitree did not open-source its VLA model for the first time until January 2026, relying entirely on external solutions before that. Founder Wang Xingxing himself admitted that the current large model capability has not yet reached the GPT-1 level. Annual R&D investment related to large models is only a quarter of that of leading domestic peers.
Third, hardware parameters cannot support industrial demands. The maximum payload of Unitree's humanoid robot arm is only 3kg, and the regular battery life is only 1-2 hours—an 8-hour factory shift means it cannot even last a quarter of it. Although UBTECH's Walker S2 also has average battery life, it can autonomously swap batteries in 3 minutes and has a 15kg payload. In industrial scenarios, these hard metrics are the threshold.
Fourth, the ROI doesn't add up. The selling price of Unitree's main models is about 167,600 RMB. Adding deployment, maintenance, and algorithm adaptation, the comprehensive cost may exceed 300,000 RMB. Meanwhile, the average annual labor cost for frontline manufacturing workers in China is 80,000 to 100,000 RMB. The common standard for factory automation equipment is a 3-year payback period—based on this calculation, Unitree robots need to replace more than 3 human workers within 3 years to break even, but their efficiency is only 50% or even lower than that of humans.
Unitree is not unaware of these issues. It raised 4.2 billion RMB in its IPO, of which 2.022 billion RMB was allocated to "intelligent robot model R&D"—using the money earned from hardware to supplement the brain.
Returning to Zibianliang's live stream. Its significance is not just an efficiency record, but proving the feasibility of another path: not piling on hardware, but relying on model capabilities to break through scenario limits. When Zibianliang outperformed Figure AI's efficiency by 45% with 30% of the cost, the technological coordinate system of the entire industry was being recalibrated.
What this list ultimately tells us is that the humanoid robot industry is diverging. On one end are "athletic stars" like Unitree, making a fortune in the scientific research market but struggling to cross the factory threshold; on the other end are "hard workers" like UBTECH, Agibot, Robot Era, and Galbot, who have already been grinding in real production lines and warehouses, with their efficiency curves continuously rising.
Backflips on the exhibition stage are spectacular, but handling and sorting in factories is where the real business lies. When the tide goes out, who is working and who is performing—this list has already provided the answer.
This article is compiled based on public network information and is for reference only. It does not constitute investment advice.