2026 World Artificial Intelligence Conference Highlights China’s Shift Toward Embodied Intelligence Industry

The 2026 World Artificial Intelligence Conference opens in Shanghai in July, with exhibition space exceeding 100,000 square meters for the first time and over 1,100 participating companies. This year’s conference witnessed a major segmentation in China’s embodied AI industry: the B2B market has clearly split, with factory-focused teams diving into industrial production lines to handle dirty, repetitive tasks, while commercial service providers flood into retail stores, restaurants, and similar settings; the C2C market divides into two camps—companion and entertainment versus household chores. Exhibitors have shifted from “showcasing technology” to “getting things done,” focusing on real-world deployment and ROI. Data collection has become a foundational industry pillar, with embodied data volumes still over ten thousand times smaller than the training corpora used for language models. Physical AI still faces three major barriers in moving from “able to demonstrate” to “able to perform”: the data wall, the representation wall, and the closed-loop wall.

Article author and source: 36Kr

Action replaces showmanship; data defines the future.

Earn money from enterprises while dreaming of creating a consumer-grade hit product.

On July 17, the 2026 World Artificial Intelligence Conference (WAIC) opened at the Shanghai World Expo Exhibition & Convention Center. This year’s exhibition area exceeded 100,000 square meters for the first time, with over 1,100 participating companies and 349 newly launched products—all record highs in the event’s history.

But more直观 than the numbers is the palpable "heat" on-site—at the entrance of the H3 Embodied Intelligence Pavilion, a long, figure-eight queue stretches out, while inside, the crowd is packed shoulder to shoulder; standing in the middle of the aisle, you can hardly remain for more than three seconds.

After visiting Hall H3, a clear impression is that the robotics exhibition has changed.

Last year’s acrobatic and dancing humanoid robots were pushed to the side—entertaining, but no longer the main attraction. Taking center stage are simulated factory assembly lines, pharmacy shelves, restaurant counters, and home living rooms. Exhibitors are no longer asking, “Isn’t this cool?” but instead saying, “See how I can do this job and save you money.”

The robotics exhibit at WAIC 2026 witnessed China’s major divergence in embodied intelligence, offering the most direct illustration of the industry’s transformation: working, diversifying, and laying foundations—three keywords corresponding to three profound shifts currently underway. Below are our firsthand observations from the exhibition floor.

Working hard is the main theme, and differentiation is the defining characteristic. Upon closer inspection, even among B2B players, the paths are vastly different. One group dives into factories, taking on the most tedious and physically demanding tasks, while another floods into supermarkets, pharmacies, and restaurants, seeking opportunities where people are closest.

The factory-style approach is straightforward: they replicate the actual production line 1:1 on the exhibition floor, letting robots "go to work" live by performing repetitive, mundane tasks.

Geek+, the market leader in warehouse and logistics robots, has launched its first general-purpose humanoid robot, Gino1, featuring full-body parallel control with coordinated hand-foot operation. It can collaborate with mobile robots and picking workstations to complete the entire picking process, reducing single-task operation time by approximately 30%.

Zhìyuán officially launched five new products: the Expedition A3 Ultra, Spirit G2 Max, Lingxi X2 EDU, Critical Point OmniHand 3 Ultra-M, and Kuituo Cycling Robot. The exhibition showcased the Spirit G2 Max in collaboration with JD Logistics, demonstrating real-world applications in warehouse handling and palletizing. A fully operational production line for tray loading and unloading was also replicated on-site. Additionally, Zhìyuán partnered with Junpu Intelligent to build a complete chip processing production line, successfully executing an extended sequence of operations—including material loading, boxing, and transportation—without any human intervention, validating its stable, autonomous performance.

Shi Zhihang moved the automotive wiring harness circular production line onto the exhibition booth. Its A1 Wiring Harness Intelligent Robotics Solution is specifically designed for end-to-end automation of wiring harness manufacturing, featuring a proprietary embodied large model capable of simulating human-like flexible operations and seamlessly integrating into existing production lines. The solution has been deployed at scale at Tianhai Electronics since June this year.

Leju Robotics demonstrated fully automated, multi-station flexible operations for pallet dismantling, material handling, and feeding—without any human intervention—operating continuously at the exhibition with performance identical to real industrial production lines.

Industrial scenarios feature high structural clarity, well-defined task boundaries, and quantifiable ROI—a consensus among factory operators. Robots don’t need to be all-purpose; excelling at one specific task at a single workstation is sufficient.

Those facing commercial service challenges encounter an entirely different set of questions, in a more open environment with more random disruptions and less room for error.

Galaxy General showcased two clearly positioned robots: the general-purpose robot Galbot G1, designed for interactive services, and the heavy-duty robot Galbot S1, tailored for industrial applications. All movements of both products are real-time driven by Galaxy’s proprietary embodied large model, AstraBrain.

G1 demonstrates versatility across both consumer and business service scenarios, covering applications such as breakfast preparation, smart retail, and guest reception; S1, on the other hand, focuses on heavier industrial applications, fully showcasing the logistics workflow from depalletizing and transporting to palletizing.

The newly upgraded AlphaBot 2 (AiBao) takes a different approach—transforming into a shop owner who prepares coffee, cocktails, and ice cream on-site, offering not just operational functionality but also the essential emotional value and consumer experience critical in B2B services.

Although the paths of these two companies differ, they share the same business logic: the more open the scenario, the greater the technical challenge, but also the more urgent the market demand and the stronger the willingness to pay. As long as reliability is proven in a specific scenario, a viable business loop can be established. This is also why most enterprises in the B2B space choose to simultaneously target both industrial and commercial service scenarios.

Notably, Fourier Robotics has launched its flagship high-load wheeled dual-arm robot, GRW, featuring a maximum stable load capacity of 16 kg, a large working space of up to 2 meters, and agile mobility—targeting high-load human-robot collaboration scenarios such as warehousing and logistics, industrial manufacturing, and elderly care. The robot is equipped with Fourier’s no-code development platform, further lowering the barrier to operation.

In addition, cultural and entertainment guidance services have become a relatively mature niche segment. During WAIC, Zhiyuan partnered with Qingtian Rental to deploy sixty various types of embodied intelligent robots across the main venue and two satellite venues, providing directional assistance in key public areas. Models such as the A3 and Lingxi X2 not only offered guidance services but also demonstrated complex movements and synchronized group dances on-site, drawing enthusiastic crowds and applause.

Peng Zhihui, co-founder of Zhì Yuán, said in an interview that the most obvious shift in the industry is from showcasing flashy demonstrations to deployment and productivity: “People are now more focused on whether it can truly enter factories and commercial service environments.”

Whether it’s a factory or a store, the paths differ, but the challenge is the same—how to secure a ticket to survival first in the commercialization process. Each company is answering in its own way, awaiting validation from the market.

The segmentation among C-end users is far less clearly defined than among B-end users. Factories purchasing robots calculate ROI, payback periods, and how much labor they can replace. In contrast, individual consumers’ decisions are much more complex: some buy because it’s “fun,” others compare prices based on “what it can do,” and still others have dual expectations of companionship and practicality—often combining all three.

From the scene, while targeting households and individual consumers, some companies position robots as “companions” or “study partners” to play chess or chat with people, selling emotional value; others position themselves as “nannies” focused on household chores, selling the value of getting things done; and still others seek a balance between companionship and task execution.

Futurion’s “Cathead” robot family has taken emotional value a step further: the desktop-grade GR Nano focuses on personal emotional companionship, subtly changing its emotional state when touched, and allows users to customize its unique personality. The GR Mini, with its cute design and human-like dancing, enhances scenarios such as exhibition guiding and elderly companionship. Although the GR Nano has not yet been released, Futurion’s strategic focus on emotional connection in the consumer market is already clear.

Qiyuan T1 and Q1 are designed to enhance family companionship and STEM education. The T1 is the world’s first transformable personal robot, capable of seamlessly switching between wheeled humanoid and quadruped forms; the Q1 features a lightweight design focused on family companionship and STEM learning. The booth is driven by interactive robot experiences, and physical flagship stores have already been opened in commercial districts.

JAKA has launched its compact humanoid robot, JAKA π ("Pi Boy"), standing just 1.22 meters tall and among the smallest humanoid robots in its class. Equipped with a "brain + cerebellum" integrated architecture, it performs and interacts with visitors at the booth. Although JAKA’s primary focus is industrial applications, the appearance of "Pi Boy" demonstrates that even industrial-focused companies cannot ignore the appeal of a cuteness-driven approach.

Zhuji Power's female Luna humanoid robot captivates audiences with its soft fabric design and fluid dance performance, featuring natural, graceful movements. It comes with an iPad control terminal that enables one-tap video-based dance learning and drag-and-drop choreography.

Weitai Power introduced a robotic dog at the JD exhibition that can autonomously accompany and chat with users without requiring a remote control.

Looking closely at these “companion” agents, their tasks have clear boundaries, interactions are controllable, and the technology is relatively mature—they don’t need to understand the complex physical world, only to respond correctly within limited rules. These scenarios are essentially “closed-ended problems,” requiring no heavy lifting; as long as they’re engaging and provide a sense of companionship, people are willing to pay.

The approach of the hardworking team is entirely different. They must enter homes to perform household chores such as folding clothes, picking up toys, and tidying rooms. For robots, these tasks are堪称“地狱难度”—fabric is soft, wrinkles are random, and the same T-shirt takes on a different shape each time it’s placed.

Furuiye’s “Embodied Home” is one of the most advanced technical implementations among the doers. The booth features a fully simulated home environment called “A Day in the Life of a Robot Butler.” When users give vague commands like “I’m thirsty,” the GR-3 can autonomously understand the need, locate the target, and deliver the item without requiring a specific location. A large screen on-site visually displays the robot’s decision-making process in real time.

The future is near—F2 is positioned as a "dedicated family管家," featuring a soft, adorable design that can play chess and supervise learning, as well as fold clothes and refill cat food. The product is now available and has been adopted by hundreds of households through paid subscriptions, following a pragmatic approach of "starting with what’s feasible and gradually expanding."

Youliqi demonstrated long-range tasks such as making pizza, drawing sugar art, and cleaning rooms, but upon closer inspection, these scenarios are more like “capability validations” and still fall short of truly addressing everyday household pain points.

A smooth demonstration on stage cannot change one reality: the home environment is an open world, while current technological capabilities remain confined to closed or semi-closed settings. Completing predefined tasks in fixed scenarios is already challenging; truly entering the diverse, individualized world of “a thousand homes, a thousand faces” still has a long way to go.

The divide between the companion camp and the productivity camp appears to be a difference in technical approach, but in reality, it reflects contrasting visions of human-machine relationships. The former views robots as companions, selling emotional connection; the latter sees them as assistants, selling labor liberation.

Under closed rules, "easy questions" versus the "hellish difficulty" of an open world reflect two different commercial paces: one scales rapidly through economies of scale, while the other advances slowly through technological breakthroughs.

Many companion platforms have already established scalable, closed-loop operations, while task-oriented platforms are still waiting for a technological inflection point. On the consumer side, the market is more willing to pay for products that are usable today rather than those promising potential in the future.

In addition to the distinctions between B2B and B2C scenarios, there was another type of booth in the exhibition hall—without dance or interaction, only head-mounted cameras, data gloves, and teleoperation stations, along with people wearing them and repeating the same motion over and over. Their task was simple: before robots can learn to work, teach them how to learn.

At the Mifeng Technology booth, attendees can experience the Mego View headset and the Mego Gripper, a claw-like handheld device. When worn, the device allows users to pick up objects in front of them; the gripper’s 200° fisheye lens, 3D haptic sensors, and nine-axis IMU simultaneously capture data, which is then reconstructed in real time on a large screen with millimeter-level precision. Mifeng’s positioning is clear: they don’t build robots—they act as “robot teachers.” By crowdsourcing human operation videos, they reduce data acquisition costs to one-third or even less of traditional teleoperation using physical robots.

Siasun Robotics follows a high-precision force-feedback teleoperation approach, offering exceptional accuracy at a high cost. With the Diana3 + Diana7G2 data acquisition system, operators hold the master arm, while the remote slave arm precisely mirrors all movements in real time, enabling millisecond-level bidirectional force feedback and capturing multimodal data including haptic, motion, and visual information.

Shi Zhihang demonstrated a data collection system built around the "Human-Centric" paradigm, featuring a self-developed high-precision SenseHub data collection system that uses a five-finger glove to fully capture hand spatial pose, finger articulation, and operational force. The core logic is: the subject of data collection is human, not robot—large models absorb human operational data before transferring it to robots.

There are far more than just these three companies offering similar data collection devices in the exhibition hall. From RoboPocket by Qiongche Intelligence, HumanData by Jianzhi Robotics, the visuo-tactile data collection gripper by Qianjue Robotics, to Looper by Chenjing Technology, data collection devices are increasingly covering more specialized dimensions—from vision to touch, from full-body motion capture to first-person perspective.

The robots on display, deployed in factories and supermarkets, appear to be performing tasks, but are simultaneously collecting valuable data from real-world environments. The work and data collection are intertwined.

Zeroth power is a prime example of this logic, centered around four key pillars: instant retail, data-driven foundations, scenario expansion, and home integration. A fully immersive space was created at the booth, where the ZERITH-H1 wheeled humanoid robot, equipped with a 3D multimodal perception system, continuously operates in front of a simulated convenience store shelf, identifying products, picking them up, sorting, and restocking—all autonomously. Every data point—whether each grasp succeeds or fails, if the force is appropriate, or if the grasp point is accurate—is fed back to the backend to drive model iteration.

The key to this logic is "physical extraction API-ization." Zeroth power transforms physical operations such as extraction, recognition, and sorting into reusable, standardized interfaces, allowing optimized interfaces to be directly invoked by more use cases.

But whether this "data-fed capability" can truly achieve generalization remains the industry's biggest unknown.

Yao Maoqing, partner at Zhixyuan, pointed out at the WAIC Embodied Intelligence Forum that to achieve scalability in physical AI, three barriers must be overcome: first, the data barrier—real-world interaction data is scarce and expensive to obtain; second, the representation barrier—a unified physical representation across tasks and scenarios has not yet been established; third, the closed-loop barrier—real-world trial and error is costly and feedback is slow.

He revealed that the scale of embodied data is more than ten thousand times smaller than the corpus used for pretraining language models, and achieving a comprehensive understanding of physical laws requires at least hundreds of millions of hours of data accumulation.

At a higher level in the data supply chain, RoboSense addresses the questions of where the data comes from and how to ensure its quality. Its second-generation all-solid-state perception platform, E2, offers three times the precision of its predecessor and achieves a point frequency in the millions, enabling robots to simultaneously generate computable and learnable data assets during movement and manipulation, forming a closed loop of “perception—data—training—iteration.”

The proliferation of data collection devices is, to some extent, compensating for a lack of intelligence with sheer effort. As Yao Maoqing noted, the transition of physical AI from “capable of demonstration” to “capable of execution” hinges on whether a generalizable, optimizable, and evolvable universal system can be established.

The head-mounted cameras, data gloves, and teleoperation stations in the exhibition hall are all doing the same thing: translating the physical world into language that models can understand. The faster and more accurately this translation happens, the sooner robots will learn to work. But until then, the entire industry must keep diligently collecting data, using the most straightforward methods, waiting for the most intelligent outcome.