Robot colleagues are starting to take on the night shifts at the logistics center.
The most frequently mentioned sentence by Xi Yue is "No one has done this before." Training humanoid robots to work from scratch is a slow process. Xingdong Jiyuan stationed at a logistics center in November last year, and it wasn't until the first quarter of this year that initial verification was achieved.
At the end of May, Figure, a US embodied intelligence unicorn valued at $39 billion, concluded a 200 - hour live - stream. This live - stream, which showed humanoid robots working at a logistics sorting center, attracted the attention of millions of people worldwide.
In Beijing, 10,000 kilometers away from Figure, Xi Yue, the co - founder of Xingdong Jiyuan, an embodied intelligence company valued at 10 billion yuan, also paid attention to this live - stream. Logistics sorting is also the area that Xingdong Jiyuan is betting on. After watching the live - stream, he said, "We entered the real - world logistics scenario earlier than Figure, faced greater task difficulties, and achieved better intelligence."
From the live - stream, the logistics center where Figure sorted express packages was clean and quiet, with a low stacking degree. The packages were transported on a circular conveyor belt. There were a total of 3 types of packages sorted by Figure: two types of cartons with different specifications and one type of soft package with different colors.
Xi Yue analyzed that this might not be a real factory production line but a POC (Proof of Concept) scenario. Currently, Xingdong Jiyuan's robots have been deployed in more than 10 logistics centers, including those of China Post and SF Express, and they can sort dozens of types of express packages. Xi Yue said that Xingdong Jiyuan is faster than Figure in terms of deployment speed. However, Xingdong Jiyuan has not yet achieved the hardware capability of Figure's single - machine continuous operation for 81 hours.
Logistics express is the commercialization scenario that Xingdong Jiyuan is betting on in 2024. Chen Jianyu, the founder of Xingdong Jiyuan, believes that this is an area that can generate high value, has technical thresholds, and can be quickly replicated and scaled up.
According to Xi Yue, after nearly a year of laboratory simulation and verification at real - world logistics centers, Xingdong Jiyuan's humanoid robots have completed the PMF (Product - Market Fit) verification and entered the large - scale deployment stage. At the beginning of this year, there were only a dozen humanoid robots at logistics centers such as China Post. By the end of May, the number had increased to over a hundred. It is expected that thousands of robots will be deployed by 2026. At these logistics centers, the humanoid robots can achieve an average efficiency of 90% of human sorters, sorting more than 1000 express packages per hour. Xi Yue expects that by the end of June, the sorting speed of the robots is expected to surpass that of humans.
The deployment in the logistics scenario has made Xingdong Jiyuan popular in the primary market. In March this year, Xingdong Jiyuan raised 1 billion yuan, with a valuation of 10 billion yuan. In April, it raised another $200 million, and many well - known institutions entered the market for the first time, including IDG and Sequoia China.
Liu Tianjie, the managing director of Huaying Capital, invested in Xingdong Jiyuan in June last year. One of the triggers for his attention and investment in Xingdong Jiyuan was that at the NVIDIA CES press conference, Huang Renxun announced six Chinese embodied intelligence cooperation companies, including Xingdong Jiyuan, Unitree, Zhiyuan, and Galaxy Universal, and Xingdong Jiyuan had the lowest valuation among them.
Qingkong Tiancheng participated in two consecutive rounds of investment in Xingdong Jiyuan. Liu Xiaojun, the chairman of Qingkong Tiancheng, believes that express sorting may be the field where humanoid robots first enter the real - world production scenario on a large scale, and Xingdong Jiyuan has overcome the model generalization task.
"Sweet Torment"
At 1 a.m. on May 28th, in a China Post logistics center branch covering nearly 6000 square meters in Longgang District, Shenzhen, six humanoid robots over 170 centimeters tall took express packages from the sorting slots, flipped them to make the waybills face up, and then placed them on the black conveyor belt beside them.
One meter away from the robots, Xiao Ke, the delivery project manager of Xingdong Jiyuan, took a sip of coffee to refresh himself. He hoped for a smooth night without any problems with the robots. From November last year to May this year, Xiao Ke and the data collection team and technical support team needed to complete the POC verification at the Shenzhen logistics branch.
Before the robots were transported from Beijing to Shenzhen, they had been tested within the company. At that time, the robots could grab 600 express packages per hour, but in the real - world branch, the robots were immediately "confused" and could only grab about 100 packages per hour.
There were many problems. First, there were changes in factors such as light and environment in the real - world operation scenario. The main working time of the humanoid robots was from 7 p.m. to 2 a.m. In the Beijing laboratory, the simulated scenario had bright lights, but in the real - world branch, the light was dim and there was reflection, making it difficult for the robots to recognize.
Second, there were various types of express packages. Xiao Ke said that he didn't know there would be so many express packages of different sizes in the world, such as kitchen knives wrapped in tape and lion - shaped stones. In the laboratory, the robots could only grab soft - packed or fixed - specification boxes and couldn't grab irregular - shaped items.
The number of express packages in the real - world branch also exceeded their expectations. In the laboratory, they could pour a maximum of a few hundred express packages into the sorting slots at a time. But in the real - world branch, there were often tens of thousands of express packages piled up like a mountain, and the robots didn't know how to grab them.
In addition to hand movements, the robot bodies also often faced challenges in the early stage. During Figure's live - stream, there was a scene where the robot suddenly scratched its head and stopped moving. Xi Yue was very familiar with this scene, which was the automatic callback action of the robot when it encountered a problem it couldn't handle.
Before officially entering the Shenzhen logistics center, Xi Yue expected the robots to achieve 60% of human efficiency. But after the actual deployment, he found that it was impossible to achieve. The customers were very anxious, frequently conducted reviews, applied pressure, and questioned the feasibility of the project deployment, believing that humanoid robots couldn't adapt to the real - world logistics scenario.
At the logistics center, Xiao Ke was responsible for recording problems every day. At first, the robots often lost connection, stopped moving, or even malfunctioned. The technical engineers had to immediately go up to handle the abnormalities, and Xiao Ke needed to explain to the customers why the robots stopped.
He worked from 9 a.m. to 1 a.m. every day. During the day, he conducted daytime equipment debugging, model training, and on - site data collection. At night, he monitored the night - shift operation of the robots and handled sudden situations such as equipment abnormalities and special express package sorting at any time. Some problems that couldn't be solved on - site needed to be iterated at the R & D center in Beijing. In the six months in Shenzhen, Xingdong Jiyuan solved dozens of problems of various sizes.
"This is a sweet torment," Xi Yue described it. During the interview, the most frequently mentioned sentence by him was "No one has done this before." Training humanoid robots to work from scratch is a slow process. Xingdong Jiyuan stationed at the logistics center in November last year, and it wasn't until the first quarter of this year that initial verification was achieved. This is of great significance to Xingdong Jiyuan. On the one hand, it can generate income. On the other hand, the team has learned the industry know - how, accumulated real - world data, and can feed it back to the embodied intelligence model.
In March this year, the sorting efficiency of Xingdong Jiyuan's humanoid robots reached 900 express packages per hour. On March 5th, Xiao Ke's team entered the second site and achieved a sorting rate of 700 - 800 packages per hour within two weeks, completing the cross - site deployment verification.
The ultimate requirement of the logistics center for Xingdong Jiyuan is that the humanoid robots should fully reach the level of humans. At the Longgang logistics center in Shenzhen, the average efficiency of manual sorting is more than 1000 packages per hour, and some skilled workers can reach a maximum of 1200 packages per hour. By the end of May, the sorting speed of the robots increased again, and the robots at the Shenzhen site could stably sort 1200 packages per hour.
Xi Yue expects that by the end of June, Xingdong Jiyuan's robots can meet or even exceed the customers' requirements.
Competing with 3 Peers for the Job
After completing the first - stage verification at the postal logistics center, Xingdong Jiyuan entered another leading logistics company. Before Xingdong Jiyuan, this logistics company had already invited three humanoid robot companies for verification. After Xingdong Jiyuan entered, it had an on - site PK with these three peers. One month later, Xingdong Jiyuan became a partner of this company.
Before trying humanoid robot sorting in logistics, this logistics company had tried everything from mechanical arms and suction cups to wheeled robots and grippers, but none of them could meet the sorting requirements. Finally, it decided to let humanoid robot companies enter the factory for a try.
There was an element of chance in this cooperation. In March this year, this logistics company had a very urgent project and was not satisfied with the work of two humanoid robot companies. After learning about it, the Xingdong Jiyuan team completed the task in two weeks and thus got the opportunity to compete with its peers.
Finally, this logistics company became the client and also an investor of Xingdong Jiyuan.
Xi Yue revealed that at the end of May, many domestic and international logistics giants hoped to cooperate with Xingdong Jiyuan, and some companies proposed to directly pay for POC deployment at their logistics centers.
During the deployment of humanoid robots, Xi Yue often went to work with the delivery team. He understood the logistics companies' needs for robots: the environment in the logistics sorting warehouse is not suitable for humans to work for a long time. The daily temperature in the Shenzhen warehouse is above 40°C, and the peak temperature in similar sites in Guangzhou exceeds 50°C. In May, Xi Yue went to the Shenzhen site, and the delivery team was sweating profusely while working in the sorting center.
There are more than 200 sorting workers at the Shenzhen logistics center where Xiao Ke works. Most of them are women over 40 years old, and they are all temporary workers with daily - paid salaries. This logistics center has always had the problem of recruiting workers because sorting is a boring, repetitive, and strenuous physical labor that often involves working day and night. The express warehouse is filled with the smell of glue from fillers and the roar of machine operations, and young people are not willing to work here.
As of the end of May, Xingdong Jiyuan's humanoid robots have been deployed at more than 10 logistics centers in 5 provinces and cities across the country to sort express packages. Different from the six - month deployment in Longgang, Shenzhen, currently, it only takes about 3 days to deploy robots at a new logistics center, and a total of 2 hours of high - quality data collection is required to complete the deployment.
Xi Yue expects that in a month, only 1 hour of high - quality data collection will be needed for deployment.
During the pilot last year, Xingdong Jiyuan delivered a dozen humanoid robots in the logistics scenario. In the first quarter of this year, dozens of robots were delivered, and by the end of May, more than a hundred robots had been delivered. Xi Yue expects that the deployment at a major client this year can exceed a thousand units.
Xingdong Jiyuan has not disclosed the specific price of its humanoid robots. The price of full - size bipedal humanoid robots of the same height on the market ranges from 200,000 yuan to 700,000 yuan. However, Xi Yue said that since the lower body of its humanoid robots is a column and does not require the structure and core components of lower limbs, the price is one - third cheaper than that of bipedal robots of the same size and also cheaper than wheeled robots.
In some logistics centers, the humanoid robots work in a "two - shift" system on a daily basis and a "three - shift" system during peak periods, achieving 24 - hour operation. Xi Yue said that this shift system can enable logistics companies to recover their costs in a relatively short time.
Dexterous Hand and Brain
The fact that Xingdong Jiyuan's humanoid robots can be deployed in the logistics industry and sort and flip express packages is closely related to the five - finger dexterous hand.
The dexterous hand looks small in size, similar to a human hand. Each finger has a driver, and there are encoders, reducers, and motors on the hand. The number of joints is not much different from that of the robot body.
Xingdong Jiyuan's first commercial order came from the dexterous hand. At the beginning of 2025, Chen Jianyu posted a WeChat Moment, mentioning the breakthroughs the team had made in dexterous hand technology and products. After seeing it, a researcher from an overseas institution contacted them and inquired about the possibility of purchasing. After using it, this institution recognized the performance and stability of the dexterous hand and made repeat purchases.
Inside Xingdong Jiyuan's Beijing company, there is a wall displaying the list of customers. Overseas customers include OpenAI, NVIDIA, Apple, Microsoft, Google, and Meta, and domestic customers include Tencent, Lenovo, Haier, and BYD. Xingdong Jiyuan has sold nearly a thousand dexterous hands in total, and an overseas giant with a market value of $4 trillion has made repeat purchases six times.
The development of the dexterous hand is related to Chen Jianyu's choice. Chen Jianyu was born in 1992 and became a doctoral supervisor at Tsinghua University at the age of 28. He studied at the Department of Precision Instruments at Tsinghua University for his undergraduate degree, which is one of the first domestic institutions to engage in research on bipedal humanoid robots. After graduating with a doctorate, he was invited by Turing Award winner Yao Qizhi to teach at the Institute for Interdisciplinary Information Sciences at Tsinghua University. Chen Jianyu has published more than 70 papers in international top - level conferences and journals in the fields of robotics, artificial intelligence, and control.
Before the humanoid robot industry became popular, Chen Jianyu was a well - known figure in the robotics field. However, in 2025, he rarely appeared in public and focused on researching dexterous hands and embodied models. He found that if robots are to work, the hand is the most important part for training the brain's ability and collecting operation data. However, the hands on the market at that time had a short lifespan and weak grasping ability.
Xi Yue said that Xingdong Jiyuan's dexterous hand made many underlying innovations in the early stage of research and development. Previously, the dexterous hand mainly used rope - driven and linkage - based solutions, but the Xingdong Jiyuan team used a direct - drive motor solution.
Liu Tianjie, the managing director of Huaying Capital, has invested in many embodied intelligence and AI companies. He is quite familiar with large - scale models and embodied model technologies, but he feels that he can't keep up when discussing academic details with Chen Jianyu.
Xingdong Jiyuan's "brain" is centered around the end - to - end embodied model ERA - 42 and also has technical systems such as world models (Ctrl - World, VLAW, etc.) and training frameworks (iRe - VLA). For the express sorting scenario, Xingdong Jiyuan has specifically optimized, distilled, and pruned the model, and also carried out a large amount of post - training reinforcement learning.
In the Beijing laboratory, the embodied model for the logistics scenario was updated once a week. After being deployed at the Shenzhen logistics center, the on - site team collected data at night, uploaded it the next morning, and the model could be updated in the morning, achieving a daily model update speed.
Delivery and After - Sales
The experience of Xingdong Jiyuan's humanoid robots being deployed in the logistics industry is a process of first having a "hammer" and then looking for a "nail".
In 2025, Chen Jianyu created a new Demo. Xi Yue thought about how to optimize the actions based on the Demo and select a suitable scenario. After a round of research, he judged that express small - item sorting was the most suitable, so he approached China Post and SF Express.
At that time, it was the period when express delivery companies were massively deploying intelligent equipment. According to data from the State Post Bureau, in 2025, L4 - level unmanned vehicles were operating in more than 300 cities across the country, and the procurement volume reached 28,000 units in 2025. In 2025, some