Reporter |
Editor | Wen Shuqi
This year, the robots at the WAIC exhibition halls have a new title: workers.
From July 17 to 20, the 2026 World Artificial Intelligence Conference (WAIC) kicked off in Shanghai. For the first time, the exhibition area surpassed 100,000 square meters, with activities spanning three venues—Expo, Zhangjiang, and West Bund—and four halls simultaneously. Over 1,100 companies participated, showcasing nearly 4,486 exhibits, including 349 global debuts, setting a record in scale.
Robots were undoubtedly the hottest and most crowded sector. This year, embodied intelligence was listed as a core conference track for the first time, with over 200 related companies exhibiting—the highest number ever. Just days before, a Ministry of Industry and Information Technology official predicted at a press conference that the annual production of humanoid robots could exceed 100,000 units in 2026.
But what’s more striking than the numbers is the shift in what exhibitors displayed. In previous WAICs, robot zones were all about flashy tricks—dancing, running, doing backflips—showing off the machines’ physical abilities. While those performances are still around this year, they’re no longer the main act.
The change first shows in how robot makers present themselves. Zhiyuan brought its factory and subway logistics warehouse setups to the venue, while Qinglang set up a coffee kiosk, laundry room, and retail store on its booth. Qiongche Intelligence demonstrated pharmacy medication retrieval, and Deep Robotics showcased flood rescue solutions… The focus shifted from individual robot capabilities to specific work scenarios.
Peng Zhihui, co-founder, president, and CTO of Zhiyuan, told NUPIAO and other media that the most obvious industry shift is embodied intelligence moving from flashy demos over the past two years to deployment and productivity, with metrics becoming more practical.
“People no longer just care if a robot can run, dance, or perform. They’re more focused on whether it can actually enter a factory or commercial service environment, work stably for long periods, and generalize across tasks,” he said. In his view, whether deployment can consistently generate data for iteration is also a key metric.
Another clear change is in the supply chain. Companies specializing in data, world models, and embodied chips and sensors set up independent booths at scale this year.
The shift in booth focus and the supply chain’s differentiation are natural outcomes of embodied intelligence entering a new phase.
Once robots leave pre-set exhibition routines and enter real factories, stores, or even homes, they face countless unfamiliar situations. This generalization ability relies on massive real-world data, which robot makers alone can rarely gather. This turns data, models, and perception into standalone businesses.
Yao Maoqing, partner, senior VP, and head of the embodied business at Zhiyuan, further explained to NUPIAO and other media that to reach the GPT moment in the physical world, the total data volume needs to far exceed the industry’s current stock. Plus, this data must closely mirror the real world, enabling robots to accurately describe physical states and predict physical laws.
Wang Xiaogang, chairman of Daxiao Robot and co-founder of SenseTime, gave NUPIAO a more specific timeline: to achieve an intelligence emergence moment for embodied intelligence around 2027, we need to accumulate tens of millions of hours of human-centric data over the next two years.
So, how did companies at this year’s WAIC address the must-solve problem of entering the real physical world? Here’s what they brought to the table.
Bringing Scenes to the Exhibition
The most intuitive takeaway from WAIC 2026 is that companies are no longer just talking technical specs. They’ve brought real work scenarios right into the exhibition halls.
Zhiyuan dubs 2026 the Year of Deployment. In its WAIC zone, the Jingling G2 Max demonstrated a warehouse deployment scenario in partnership with JD Logistics—the first time a humanoid robot has been deployed in a real warehouse operation. The robot can lift 18 kg per arm and work 24/7 on tasks like moving and stacking pallets. They also recreated a semiconductor tray handling production line on-site.
Beyond the venue, Zhiyuan teamed up with robot rental platform Qingtian Rent to deploy 60 robots for wayfinding services in public areas of the main and sub-venues—the industry’s first “deployment mode” public service for embodied intelligence.
In interviews, Peng Zhihui said embodied intelligence will first land in business scenarios where tech maturity matches demand, with factories being Zhiyuan’s primary entry point. “We’ve already deployed robots on real production lines in some manufacturing companies in Jiangxi, including large 3C electronics factories, where they’re working regularly,” he added. Besides factories, commercial service environments like shopping guides and tours are another key focus.
If Zhiyuan brought factory-like scenarios into the booth, Qinglang took a service-industry approach. It exhibited in an “embodied community” format, with humanoid robots operating fully autonomously—no remote control—across four scenarios: a coffee kiosk, dessert shop, retail store, and laundry room.
Wan Bin, COO of Qinglang Intelligence, told NUPIAO that working in real scenarios presents massive engineering challenges compared to demo products, but the industry’s progress is marked by more players investing in practical, useful solutions.
Retail stores are another scenario being validated. At WAIC, Qiongche Intelligence demonstrated pharmacy medication retrieval. Its robot, powered by the Qiongche embodied large model, identified and grabbed items from over 3,000 SKUs, needing only about 2.5 square meters to operate without store modifications. The solution has already been validated in chain pharmacies in places like Shenyang.
Beyond standardized factories and retail stores, hazardous and high-risk special operation zones are naturally suited for robots.
Deep Robotics highlighted emergency rescue as one of its key scenarios this year. Qiao Xin, Deep Robotics’ solution director, told NUPIAO that for rescue in harsh environments like landslides or typhoons, robot dog combos can now step in. For instance, a robot dog can advance 3 to 5 km to scout dangers and send back video, completing site surveys and data communication before human rescuers arrive.
During the conference, Deep Robotics showcased standardized solutions for six industries: power, emergency, public security, firefighting, forestry and grassland, and education.
For complex rescue needs, Qiao Xin believes the future lies in “combo operations,” with robot dogs excelling at danger scouting and personnel search, while larger robots handle heavy lifting like clearing collapsed buildings. Product forms will move toward diverse combinations.
At Unitree’s booth, the spotlight went to the global debut of the GD01, the world’s first manned deformable mech. Xia Geng, Unitree’s marketing manager, told NUPIAO that the GD01 stands about 3 meters tall, weighs around 500 kg with a person onboard, and can switch between bipedal and quadrupedal modes. It’s designed to replace humans in dangerous scenarios. Priced at 3.9 million yuan since its May launch, it’s a game-changer.
Beijing Humanoid targets dirty, heavy, and high-risk special operations. Leveraging its self-developed “Embodied Tiangong” humanoid robot tech system and the “Huisi Kaiwu” general-purpose embodied intelligence platform, it offers packaged solutions for commercial services, logistics sorting, electronics manufacturing, energy inspections, and more.
Fourier Intelligence brought its first full-stack embodied intelligence tech demo for home companionship—the “Fourier Embodied Home” solution. The GR-3 acts as a smart home assistant, handling tasks from security monitoring and item delivery to companionship and guest reception.
At the same time, as technology generalizes and costs drop, robots are starting to enter personal and consumer markets. But compared to factories and retail stores, companies are taking very different approaches.
Shangwei New Materials’ Qiyuan Robot T1 made its debut at WAIC, focusing on form. This personal robot can switch between wheeled humanoid and quadruped modes, moving quietly indoors and tackling grass and stairs outdoors. It can even sync with an action camera for follow shots.
Songyan Power bets on price and emotion. Its thousand-yuan Bumi robot danced in groups at the booth, while the new bionic robot Xiaoyue interacted with visitors through eye contact and facial expressions.
But making robots truly versatile across these scenarios takes more than hardware. Whether they’re screwing in factories or being home butlers, all companies face a common bottleneck: robots must adapt to unknown real environments, which heavily depends on model generalization, and that’s built on massive data.
Solving this pain point is the mission of another emerging group at this year’s WAIC.
Data and World Models Become Independent Businesses
This year at WAIC, beyond the flashy robot makers, companies focused on world model R&D, data collection, and core component supply set up their own booths at scale. This marks a rapid specialization and clustering of the embodied intelligence supply chain.
Wang Xiaogang told NUPIAO and other media that what the industry truly lacks is generalization ability in unknown scenarios. Many impressive demos still work only on pre-set tasks. Once robots enter real production and life scenarios, they encounter countless new situations.
Wang Cong, CEO of Dijia Robot, told NUPIAO and other media that this year’s improvement in algorithm generalization isn’t due to a single breakthrough technology. Instead, it’s because the industry has done more solid foundational work in data accumulation, alignment, and governance.
He used autonomous driving as an analogy. “When the industry started collecting data, there were no mass-produced cars. No one knew how to mount cameras or how many radars to use. The first batch of data was mostly wasted.” In his view, embodied intelligence will go through the same phase. “After wasting a few rounds of data and taking a few detours, a viable path will naturally emerge.”
In other words, the ceiling of compute and algorithm power is now largely determined by the scale of high-quality data.
Yao Maoqing said the industry is waiting for a GPT moment, but it requires a systematic, society-wide effort. He revealed that Zhiyuan’s data accumulation target this year is close to 10 million hours, with the goal of collecting skills data from all walks of life for AI. “Only real-world data can best describe physical laws.”
The debate between VLA (Vision-Language-Action) and world model tech routes remains a hot topic.
Wang Xiaogang believes they’ll coexist in the short term. When world models aren’t fully mature, VLA learns faster for high-certainty tasks and executes more reliably, making it a good supplement. World models, on the other hand, are better at understanding and breaking down long, complex tasks.
But he thinks as data scales up, world models will gradually absorb VLA’s capabilities, eventually leading to a unified base model.
This route evolution directly impacts the “logical reasoning” ceiling of robot brains.
Guo Song, associate head of the Department of Computer Science and Engineering at Hong Kong University of Science and Technology, told NUPIAO that compared to autonomous driving, humanoid robots must interact with complex physical worlds. They not only need a cerebellum for control loops but also a brain with long-range logical reasoning to understand human intentions and complete sequential tasks like cleaning a desk.
He sees the future as a merger of the simulator (world model) and the planner (VLA), integrating simulation capabilities into VLA’s action abilities to bridge the intelligence transmission chain.
Liang Quan, VP of marketing and ecosystem at Arm China, added from a chip deployment perspective that VLA models are currently structurally stable and have closed loops in specific scenarios, making them the most practical choice for direct deployment on domestic chips. That’s why Arm China is deepening its collaboration with HKUST on VLA model development and deployment.
His view is that, compared to the still-rapidly evolving world models, VLA can help the supply chain build an ecosystem and commercial positive cycle faster.
Additionally, Yao Maoqing believes there’s too much marketing jargon in the industry. Everyone likes to slap fancy labels on their tech. VLA and world models are just descriptions of product forms, not tech essence. Peel away the terms, and what robots truly need is a multimodal large model for the physical world—one that can take in text, images, video, and force data, and output actions, plans, and predictions of future scenes.
He argues the real dividing line is that this model’s structure and training must be designed from scratch around tens of millions of hours of embodied native data, not adapted from entertainment-oriented video generation models or image-text Q&A models.
The hunger for high-quality spatial data has directly birthed a new infrastructure ecosystem. To fill the data gap, from front-end data collection hardware to back-end model training architectures, players are racing to claim their spots in this new data infrastructure. New entrants are carving out different niches in the data chain.
On the model side, X-Era Lab showcased a robot that can tie its own shoelaces, demonstrating its native world action base model’s ability to handle soft objects and long tasks. The company revealed it has accumulated tens of millions of hours of 4D data, pre-training the model with massive real-world interaction data to give robots stronger general operation capabilities.
On the collection end, JD.com approached it from a data infrastructure angle, showcasing its self-developed JoyEgoCam head-mounted data collection device for first-person work data. It also open-sourced the industry’s largest human first-person dataset, EgoLive, aiming to build a full-chain embodied data infrastructure.
Hardware suppliers turned sensors into data entry points. Seyond (RoboSense) launched the second-generation all-solid-state perception platform E2, using high-precision digital LiDAR to simultaneously generate high-quality 3D spatial data during robot movement and operation, essentially turning the underlying sensor into a data entry point for physical AI.
The perception puzzle isn’t just about vision. Yimu Technology, exhibiting for the first time this year, focuses on vision-based tactile technology. Li Zhiqiang, founder and CEO of Yimu Technology, told NUPIAO that currently, the scarcity of tactile data and weak modal alignment remain the industry’s main bottlenecks. But over the past six months, consensus on the value of touch has been forming rapidly.
In his view, true breakthroughs still depend on specific scenarios. High-quality data loops must be built through real-world deployment to complete the “final piece of the puzzle” for embodied intelligence.
This specialization extends to the interaction layer. Mosi Intelligence, also a first-time exhibitor, showcased MOSS-VL-Realtime, a real-time video understanding model that answers questions with continuous video input, plus a multi-speaker long audio transcription model. Xin Yan Robot’s Bubbo 1 uses multi-sensor fusion to perceive a user’s voice, movement, and distance, then decides whether to respond and how, based on ongoing memory.
Some companies are trying to reassemble these specialized capabilities into a general-purpose base. Mech-Mind, already deployed at scale in automotive and 3C electronics factories, uses a “one brain, multiple forms” approach, configuring perception, understanding, planning, and execution abilities into humanoid robots, industrial arms, and mobile robots. It publicly showed its new Mech-Hand multi-fingered dexterous hand for the first time. The company says it has deployed over 27,000 units globally.
As the supply chain breaks down into finer pieces, a subtle symbiotic and competitive relationship is forming between robot makers and these new players in data, models, and perception.
Wan Bin told NUPIAO that some companies can build their own data and model loops, while others need an ecosystem approach to fill gaps. “Division of labor will lower the entry barrier for more participants. In the future, we’ll see both integrated software-hardware paths and open integration paths. But the ultimate test has only three criteria: Is the product useful? Can it run reliably? Does the math work?”
Qiao Xin believes the future industry structure will be “each has their own specialty.” Data collection and model training will be handled by specialized companies, while robot makers focus on refining their hardware and industry applications, eventually forming a thriving upstream and downstream ecosystem.
When asked when this ecosystem will spark a true tech explosion, Qiao Xin said the industry is at a critical point of cerebellum-cerebrum coordination. “It might take 3 to 10 years to break through this paper-thin barrier, but given China’s current iteration speed, we might have a very mature answer within 3 to 5 years.”
At the end of the day, whether it’s diving deep into scenarios or obsessing over data models, the ultimate goal of embodied intelligence at this stage is no longer just tech validation.
Peng Zhihui told NUPIAO and other media that practitioners of embodied intelligence must not only maintain saturated investment in cutting-edge R&D but also stick to tech for good and people-centric values.
“We need to apply our products and algorithms to real scenarios that serve society. We must look up at the stars and venture into uncharted territories, but also keep our feet on the ground and work hard,” Peng said. “The ultimate destination of technology is always to serve society.”
