At the 2026 China International Fair for Trade in Services, the "Physical AI and Embodied Intelligent Robotics Innovation Ecosystem Exchange" took place in Beijing on September 11. Zhao Yuli, Chief Strategy Officer of Beijing Galaxy General Robotics Co., Ltd., delivered a keynote speech. Below is the full transcript of her presentation.
I'm honored to join this forum and take this opportunity to share Galaxy General's latest progress. Today's theme is "From Championship Glory to Industrial Implementation." The recently concluded Beijing Humanoid Robot Sports Meet has drawn widespread attention, particularly in track and field where robots have already broken human world records. However, beyond competitive events, there's another arena deserving greater focus—the scenario competition. During this year's preparation, the organizers, Beijing's Economy and Information Technology Bureau and the General Administration of Sport, fully incorporated suggestions from all parties and set exceptionally high standards for the contests. Their core intention was to present robots with challenges genuinely comparable to human work capabilities.
Due to the extreme difficulty of the tasks, some teams were still permitted to use teleoperation methods this year, while those insisting on fully autonomous operation received higher incentives. Galaxy General was among the few companies that chose to compete using embodied intelligence large models. We firmly believe the industry truly needs such challenging tracks, using real and complex tasks as the litmus test for robots' capabilities and application readiness. During the opening ceremony, we were also among the few teams entering as "working robots," and uniquely, the only squad demonstrating a robot autonomously playing tennis on-site. We engaged in a rally with tennis world champion Zheng Jie, who spared no effort, hitting angled shots from both the front and back courts—yet the robot successfully returned every ball. Before an audience of 30,000, we completed over 100 consecutive rallies.
At this sports meet, we participated in three live competition segments, also the most challenging events overall: home, dining, and retail scenarios. We secured gold medals in all three, each achieved fully autonomously and at the fastest speed. The home scenario is particularly noteworthy. Home service has always been the ultimate dream for humanoid robots entering end-user applications—everyone hopes a robot can one day assist with countless tedious daily tasks at home. This competition's home scenario took place in a real environment, where no one except the two technical operators could enter the competition area. The robot faced non-fixed, non-standardized tasks, needing to autonomously understand and execute work in a live setting. For instance, encountering different types of clothes or pants, the robot would first self-learn how to shake out the garments before organizing them. Throughout the fully autonomous process, tasks could change at any moment—like when a doorbell rang, the robot had to respond promptly, go to the entrance to retrieve a package, and place it in a designated spot.
The dining scenario was equally demanding. Hot food is easily damaged during grasping, while tasks like operating a microwave or opening and closing cabinets involve numerous fine motor actions, placing high demands on the robot's perception, manipulation, and control abilities. Retail represents the application domain Galaxy General has consistently championed. To date, we've deployed 200 Galaxy Capsules nationwide. This competition further validated our latest technology on-site: the robot could not only pick items but also restock, organize shelves, and automate the entire workflow. For example, when products on shelves were placed incorrectly, the robot could identify the issue purely through visual perception and autonomously adjust items to their proper positions. Critically, these misplacements were randomly set by referees before the robot entered, meaning each robot faced completely different problems upon entering the arena. Ultimately, Galaxy General swept gold, silver, and bronze medals in the retail scenario competition.
Throughout the event, our technical team insisted on using embodied large models for fully autonomous operation. The industry offers various robotic operation technical routes; traditional planning and control methods can accomplish some tasks, but Galaxy General chose to complete the entire competition using the Galaxy Brain embodied large model. As a result, all three segments were finished fully autonomously, at record speed, and with near-perfect scores. In tennis, we also achieved another breakthrough in human-robot play capability. When we first unveiled robotic tennis technology in March, robots could already sustain over a dozen consecutive rallies. This time, one robot played over 100 consecutive rallies against two human players. Looking at it today, the real bottleneck limiting continuous play is no longer the algorithm but battery life—theoretically, it can keep playing until the power runs out. During this process, the robot showcased extreme ball-saving capabilities, running continuously across the court, and even if it fell while saving a shot, it could quickly get up autonomously and continue the match. Witnessed by 30,000 live spectators, we completed over 130 consecutive rallies, setting a new world record. And this involved one robot versus two human players, a higher degree of difficulty. Interestingly, this video wasn't deliberately promoted overseas, yet it drew attention and amazement from global tech figures like Elon Musk and Andrew Kang, with Musk calling the performance "insane." This signifies that China's embodied intelligence is gradually building its own technological brand on the world stage. We've named this moment "AstraTennis." A decade ago, AlphaGo defeated Lee Sedol, marking AI's breakthrough over top human performance in Go. Now, ten years later, we aim through AstraTennis to enable robots to genuinely compete with humans at high-level athletic events, targeting next year to reach a competitive level comparable to human sporting competitions.
But let's return to the industry focus. Galaxy General's ultimate goal remains empowering industrialization through embodied intelligence to achieve genuine real-world applications. We believe the inflection point for the industry has arrived, and embodied intelligence is entering a phase of rapid growth in scaled applications. In the past, when discussing large-scale robotic adoption, the most attention went to the "brain"—the robot's intelligence. However, Galaxy General isn't just building the "brain." Our team has been rooted in this field for a decade, a true band of industry veterans. Over ten years ago, we began developing humanoid robots, accumulating mature expertise in mass production, reliability, and supply chain management. Therefore, Galaxy General is truly building a "trinity" system. First, designing robot products and application logic based on real-world use cases. Second, leveraging the data loop generated from actual applications to continuously drive embodied large models, achieving true generalization and versatility. Third, in terms of overall hardware reliability and mass production capacity, we've completed over two years of real-world application validation.
Behind this system lies Galaxy General's complete full-stack technical capabilities. This year, many overseas friends jokingly refer to Galaxy General as "China's Tesla." The robotics industry indeed shares many similarities with the new energy vehicle sector. If hardware represents the vehicle's body, then the embodied foundational large model serves as the "engine." One of Galaxy General's strongest areas is data. We continue to build a high-quality data foundation, creating an embodied intelligence data infrastructure exceeding 20 billion data points through a multi-dimensional, five-tier diverse data structure, supporting the sustained advancement of this emerging track. In May, Galaxy General released the world action model AstraBrain WAM, a concept first proposed by our team at ICCV in March 2025, making it the world's first world action model framework. The world action model aims to further integrate the strengths of two technical routes—VLA and world models—allowing robots not only to understand and predict their environment but also to comprehend their own actions within it and execute precise physical interactions. AstraBrain WAM also addresses another critical issue: how to more effectively utilize data of varying quality and from different sources. Through this approach, we can expand data usage scope, enhance data effectiveness, and ultimately transform data "waste" into data "fuel"—a technically significant breakthrough.
Driven by this technological system, we've further validated the scaling law in embodied intelligence. This explains why an increasing number of experts and media are asking: Could the "ChatGPT moment" for embodied intelligence arrive within the next two to three years? Our answer is: very likely. During this year's World Artificial Intelligence Conference, we demonstrated the Galaxy Brain's model capabilities up close. We had the robot perform an ultra-long-horizon "make breakfast" task, running continuously for 8 hours throughout the day. This was one of the longest-horizon task demonstrations at WAIC. The entire task comprised multiple complex subtasks while remaining open to interruptions and challenges from live audiences. For instance, viewers could move bread or shift juice cups, requiring the robot to re-perceive its environment, adjust its motion trajectory, and ultimately close the task loop. What was showcased throughout was the embodied large model's robustness in complex real-world settings.
At this year's World Robot Conference, we also launched the bipedal robot Galaxy Xingzai. A key reason for introducing Galaxy Xingzai is to continuously push the boundaries of whole-body motion control technology and explore the limits of whole-body control algorithms. One of its standout features is the ability to "see, think, and learn simultaneously." It can watch human performances to learn dance moves, demonstrating highly human-like athletic abilities. Today, Galaxy General has established scaled partnerships with industry leaders such as Germany's Bosch, Siemens, and domestically with CATL. During Trump's visit to China this May, when accompanying media came to the country, among the robot companies he visited, only one was Galaxy General. At that time, a US-based streamer with tens of millions of followers came to China and asked if he could visit Galaxy General's showroom. I told him we didn't have one. Galaxy General is one of the few robot companies without a dedicated showroom—the reason is simple: office space in Haidian is extremely valuable, and we prefer to allocate more space to our research personnel. So, I took him directly to real-world application sites instead. I told him, rather than seeing robots in a showroom, it's better to visit our pharmacies and convenience stores to see how robots actually work in authentic environments. He toured every scenario and was astonished by each. In one setting, he placed an order in English: "May I order sausage please?" Our Galbot confidently replied in Chinese, "Grilled sausage, 8.88 yuan." Some asked why the robot answered in Chinese to a foreigner. I said, that's our cultural confidence.
Here's our smart pharmacy. In September 2024, Galaxy General partnered with Meituan to open the first smart pharmacy. China's e-commerce and instant retail are already highly prevalent; many people choose to order medicine online with 20-minute door-to-door delivery. Behind this efficient service lies a massive labor force. In such pharmacies, goods are typically densely packed, and many warehouses are located in basements without windows, requiring grueling 24-hour shifts in three rotations. For operators, another practical issue is that workers often struggle to stay in such high-intensity, cramped environments long-term, leading to high turnover rates. Therefore, these are the "tough, dirty, tiring" jobs that desperately need humanoid robots as assistants. For pharmacists, their true responsibility should be providing professional medication guidance to patients, not spending hours each day organizing shelves, picking items, and packing orders. Today, these tedious tasks are handled by Galaxy General robots. We've now implemented applications in nearly 100 smart pharmacy locations across over 33 cities. Our first store opened in September 2024, and one robot within it has been running continuously for 24 months. That means we've validated two years of normalized robotic operations.
This is our Galaxy Capsule, which many of you are familiar with. It can now be seen at locations like the Summer Palace, Wangfujing, and Badaling, deployed across 200 sites in 50 cities nationwide. This is our heavy-load robot, the Galbot S1, the world's first robot achieving a 50-kilogram payload capacity. Currently, the Galbot S1 has entered Siemens' factory operations. CCTV Finance has also recently highlighted our collaboration with Siemens, as it genuinely addresses numerous tedious, repetitive tasks in factories. For example, material handling tasks like this one previously required workers to make dozens of round trips within the factory, walking hundreds of meters cumulatively, consuming significant manpower and time. Through our robot, the entire transport process can be automated. During this process, from navigation to target recognition and environmental perception, everything is autonomously handled by the embodied model. Additionally, we've thoroughly designed human-robot collaboration safety. Working in the environment, if any worker passes by, the robot can identify them through 360-degree environmental perception, stop immediately, and maintain a safe working space.
Furthermore, we've achieved implementation at CATL's factory, becoming the first robot to operate autonomously and consistently on CATL's smart production line. CATL has extremely high requirements for production safety and task completion rates, demanding 100% completion. Previously, they evaluated multiple humanoid robot companies, but none met their stringent standards. This is because CATL's automation level is already extremely high, boasting four global lighthouse factories. What remains for manual labor are often more tedious, complex, and high-risk tasks involving precision components like battery cells. These are precisely the scenarios where humanoid robots should truly deliver value. Currently, the Galbot S1 is undertaking relevant work at CATL's factory, maintaining normalized, continuous operation.
These are the latest updates and sharing from Galaxy General. Thank you all.
Comments