How to avoid robots becoming mere decoration after dense deployment: AgiBot discloses operating accounts and deployment data

As night falls, the Hankou Riverside Promenade in Wuhan is bustling with people. Recently, the full-size humanoid robot "Yuanzheng A3 Ultra," standing 1.74 meters tall, set out from Shanghai and arrived at the "Jiangyu" bridge-pier film camp on the Hankou Riverside Promenade to officially launch its smart retail business. This is a smart retail warehouse built by AgiBot in partnership with Wuhan Culture and Tourism Group and Zhongyun Weixun. According to company disclosures, as of October 25, the robot had been running stably for 24 days, completed a cumulative 1,253 orders, and achieved a grasping success rate of over 98%; during the peak visitor flow of the most recent weekend, it even set a record of over 100 orders in a single day with a 100% grasping success rate. In less than a month, AgiBot's robots have successively entered scenarios such as Chimelong Paradise and factory production lines, with service applications including retail and medical guidance rolling out one after another. Deployment has accelerated, but can the robots be used continuously, and how can they avoid becoming "mere decorations once on duty"? From the Hankou Riverside Promenade to dense deployment across multiple scenarios Identifying products, grasping snacks and drinks, and delivering them to visitors—a process originally completed by sales staff now features robots. Unlike traditional vending machines that can only dispense goods via buttons, the Yuanzheng A3 Ultra uses a full-size bipedal humanoid body to autonomously complete the entire process of product recognition, precise grasping, and delivery, covering more than ten kinds of snacks and drinks, and engaging in human-robot interaction with visitors, directly facing the test of heavy nighttime crowds on the riverside promenade. As a commercial-grade full-size humanoid robot, the Yuanzheng A3 Ultra is 174 centimeters tall. The entire machine is equipped with a 360° fusion perception system combining LiDAR and multiple cameras, paired with 700 TOPS of computing power and a self-developed three-layer heterogeneous computing architecture, enabling autonomous mapping and dynamic obstacle avoidance and adapting to open outdoor sites; it comes standard with dexterous hands featuring up to 20 degrees of freedom with tactile sensing, enabling force-controlled fine grasping, and can stably pick up and deliver items ranging from light small-packaged snacks to heavy canned drinks. Relying on UWB and RTK centimeter-level indoor-outdoor fusion positioning, 8-hour commercial battery life, and support for autonomous charging and a fourfold active safety protection system, the device can operate unattended for long periods in the riverside scenic area, reducing on-site operation and maintenance pressure. The smart retail warehouse at the Wuhan riverside is not an isolated case of AgiBot's recent exploration of commercial applications. On September 24, the embodied intelligence-themed amusement park project jointly developed by AgiBot and Chimelong Group landed in Hengqin, Zhuhai, with more than 300 robots entering Chimelong Spaceship Park, covering seven major scenarios including cultural entertainment performances, science popularization and study tours, guided tours and shopping guidance, service retail, intelligent companion tours, hotel services, and sports competitions. At the same time, robots are also extending into production scenarios such as factories. AgiBot's 15,000th robot was previously delivered to the Longcheer production line to provide operational capabilities for industrial manufacturing scenarios; its 20,000th robot was delivered to Chimelong Group on September 24 to explore cultural tourism service scenarios. In the medical service field, applications such as medical guidance are also expanding, with robots taking on auxiliary work such as triage consultation and route guidance to help patients find departments faster and understand the consultation process. In less than half a year, the mass production figures disclosed by the company rose from 10,000 to 20,000 units, and the large-scale deployment of embodied intelligence is accelerating. AgiBot robots deployed in Kuanzhai Alley: audiences watch the excitement, but it may not be truly lively Retail, cultural tourism, manufacturing, and healthcare appear to be several different tracks, but they are important signals of technological development: robots no longer only need to complete one impressive action; they must also work continuously in specific positions and truly solve problems. However, the broader the scenarios are rolled out, the more complex the test becomes. In factories, robots need to complete operations stably; in hospitals, route guidance and information accuracy are crucial; in scenic areas, beyond completing tasks, visitors must also be willing to participate. Different scenarios have different requirements for robots, and being able to complete a task in one place does not mean it can be copied elsewhere. Especially in cultural tourism scenarios, robots can easily attract visitors because of their novel appearance, but they may also lose attention after the excitement fades. Recently, at the 2026 Global Developer Pioneer Conference and International Embodied Intelligence Skills Competition held in Zhangjiang, Shanghai, what is being compared has also quietly changed: robots on the competition floor no longer compete over who can do backflips or walk a tightrope; instead, real working conditions such as energy inspections, supermarket retail, and handling in confined spaces are brought directly onto the stage, letting a thousand participating robots face real "work assessments." This situation has sparked the same heated discussion in the industry: no matter how many spectators gather, if it cannot be converted into effective work, the popularity will ultimately be difficult to sustain. In actual operations at Chimelong in Zhuhai, Wang Chuang, AgiBot partner, senior vice president, and president of the General Business Division, observed an intriguing phenomenon: the projects visitors are willing to queue up to experience are not necessarily the ones with the highest technical difficulty. "Projects with strong interactivity, such as rock-paper-scissors, three-in-a-row, and table tennis, which children can participate in and the whole family can enjoy together, are actually more popular," Wang Chuang said. The team will continue to monitor the actual usage rate of each location and each robot, as well as visitor feedback on the experience. If visitors are found to show little interest in a certain location, continuous optimization is needed, and even equipment or content may need to be updated. In addition to attracting visitor flow, whether robots can remain on the job over the long term also involves a specific operational calculation. Beyond hardware depreciation and electricity, human-robot collaboration, networking, model token calls, and spare parts wear and tear all directly affect overall costs. Xiong Yan, AgiBot partner, co-president, and president of the Central R&D Department, revealed that in the initial stage of the Chimelong project, the ratio of on-site personnel to robots was about 1 to 3. "This means that putting robots on duty does not mean human employees can immediately leave the scene. In the initial deployment stage, robots still need human assistance, and there must be people on site to handle equipment anomalies, operations and maintenance, safety, and other emergencies. Only as autonomous operation capabilities improve and the operations and maintenance system matures can the human input behind each robot gradually decrease."Currently, companies hope to gradually raise this ratio to 1 to 10, with an ideal target of 1 to 20. Fishing for data through continuous trial and error in real-world scenarios. In the first half of 2026, global humanoid robot shipments approached 25,000 units, of which the Chinese humanoid robot market maintained rapid growth, with shipments exceeding 19,000 units. The consensus reached by all parties at the International Embodied Intelligence Skills Competition also lies in this: the industry is accelerating from small-batch trials to large-scale application. "Building it" has been preliminarily solved, while "using it well" has only just entered the deep-water zone. In response to the pain point that robots "perform well in the lab but are difficult to implement in practice," the industry's toolchain is accelerating efforts to fill the gap. On October 23, AgiBot released the one-stop full-process embodied intelligence operation platform GenieStudio 2.0 at the "AIWAVE" event, covering the entire lifecycle of "collection-training-testing-deployment-programming-management," bridging the key shortfall from model capabilities to real on-the-job work, and compressing the zero-to-one deployment validation of a brand-new operation scenario to as fast as 7 days. However, frontline R&D personnel remain cool-headed about this wave of implementation. "For robots to truly achieve what the public considers complete independence from on-site debugging and work as smoothly as real people, at least 3 to 5 years of technological accumulation are needed," Xiong Yan said bluntly. "But the industry cannot wait until the technology is absolutely perfect before entering the field. Plunging early into thousands of industries to 'build while modifying' is precisely the necessary path for technology to mature." The core value of robots densely entering all walks of life to "work" lies precisely in fishing for data from real scenarios, especially those unexpected errors. For example, in front of a hospital guidance desk, a robot faces thousands of inquiries every day. Among them, dialects it cannot understand and obscure routes it cannot answer are all recorded and sent back to the cloud. These deviations and errors encountered on site are valid data that cannot be simulated or obtained in the laboratory, and they are also indispensable real nourishment for the continuous iteration of algorithm models.