No more showing off! WAIC robots focus on practical applications.Article author and source: Lei Technology
On July 17, the 9th World Artificial Intelligence Conference (WAIC) officially opened at the Shanghai World Expo Center. As an annual flagship event in the AI industry, this year’s conference once again brought together domestic and international large model companies, AI applications, hardware manufacturers, and robotics firms. Leitech (ID: leitech), through its AI-focused new media platform “Leitech AGI (leikejiagi),” sent a reporting team to cover the event live in Shanghai.
Image source: Lei Technology production
At this year's WAIC, it's clear that the humanoid robotics industry has completely changed its tone.
In previous years, robots at major exhibitions were primarily designed for entertainment—performing backflips, dancing, boxing, and more. This year at WAIC, robots simultaneously served as both products and staff members. Various humanoid robots were deployed throughout the venue’s corners, specifically guiding visitors and managing unmanned retail kiosks. This marks the first time in the industry that full-sized humanoid robots have been widely deployed for commercial use in public areas of a large-scale exhibition.
The flashy performances have stepped aside, and mass production and practical implementation have become the core theme of this year.
The industry is opening up, and robot manufacturers each have their own strategies.
Zhì Yuán: The star attraction of WAIC 2026, defining the year of deployment
Agibot is one of the few companies in the humanoid robotics industry that has delivered tangible results in both technology and deployment. It is not only the most popular at exhibitions but also the fastest in real-world applications.
Zhiyuan has set the theme for this year's WAIC as "The Year of Deployment," and has introduced multiple deployment-focused products and scenarios, including the Yuanzheng A3 Ultra, Spirit G2 Max, Lingxi X2 EDU version, Critical Point OmniHand 3 Ultra-M, and Coolbot's world-first cycling robot.
The Expedition A3 Ultra is the only humanoid robot among the top 10 exhibit highlights at this year's WAIC. Standing 174 cm tall with over 50 degrees of freedom, it features dual arms capable of carrying 38 kg and is equipped with the latest dexterous hands integrated with an end-to-end VLA large model, enabling stable execution of complex operational tasks in real-world industrial environments.
Visit the deployment area of the Zhiyuan booth, where Zhiyuan has recreated a real-world scenario of tray loading and unloading on a production line. The精灵G2 robot autonomously identifies, picks up, and transports chip trays, precisely completing the entire loading and unloading process to meet the cleanroom and high-repetition requirements of semiconductor manufacturing facilities. On-site, a complete chip processing line by Junpu Intelligent demonstrates the robot’s ability to perform an extended sequence of tasks, including continuous chip loading, packaging of finished products, and full-carton transportation.
Image source: Lei Technology production
In the interactive deployment area, Zhìyuán also brought real-world scenarios from the Guangzhou Metro to the WAIC site, where the Spirit G2 can autonomously perform routine tasks such as security guidance and customer service. A miniature version of the tourist service center at Chengdu’s Kuanzhai Alley has also been unveiled at Zhìyuán’s exhibition booth.
For inspection scenarios, its subsidiary brand Zhiyuan Kuituo has introduced two proprietary core engines—“Lingyue” and “Lingtou”—to address key limitations in quadruped robots. The Lingyue All-Terrain Motion Engine, built on a perception-plus-motion-control dual-mode architecture, enables seamless navigation through complex industrial terrains. The Lingtou Full-Range Navigation Engine solves the common issue of positional drift in inspection robots, and when paired with automatic charging, supports months of unattended operation. This solution has already been deployed in inspection projects across power, chemical, and campus security sectors.
On the technical foundation, AgiGO-2’s proprietary embodied foundation model serves as the decision-making brain, introducing Action Thought Chain and Asynchronous Dual-System architectures to seamlessly connect planning with execution. The world model GE-2 ranked #1 overall in the 2026 World Arena World Model competition, evolving from “world simulation” to “world learning.”
Outside the exhibition area, Zhìyuán has deployed 60 robots to serve the WAIC conference, managed by its RaaS platform, Qingtian Rent.
Sixty robots operated continuously for several days, with shift changes, charging, and fault handling handled entirely without human intervention. Deploying all sixty full-sized humanoid robots to run nonstop in a public area at a large exhibition was itself a stress test.
Sweet Potato: A 560 TOPS computing foundation that significantly shortens development cycles.
What’s interesting about Digu is that it doesn’t build robots—while everyone is focused on humanoid robots themselves, it chose a different path: becoming utilities like water, electricity, and gas.
Therefore, the content displayed by Digu might appear quite technical to ordinary users.
Among them, Xuri 600 is a high-performance SoC designed for embodied intelligent robots, featuring computing units such as CPU, DSP, GPU, NPU, and ISP, specifically optimized for intelligent tasks including robotics, manipulation, planning, perception, and control, delivering powerful computing, high reliability, and low-latency responses for humanoid robots and embodied intelligent systems.
Image source: Digua Robotics
Specifically, the旭日S600 offers 560 TOPS (INT8) computing power, powered by a 4-core BPU Nash for efficient model inference. The 18-core A78AE CPU features a high-performance CPU cluster for flexible task scheduling, while the 6-core R52+MCU is designed for real-time control scenarios, emphasizing stability and reliability. Additionally, it employs 256-bit LPDDR5x high-bandwidth memory with a bandwidth of up to 204.8 GB/s, providing robust high-bandwidth data support for edge computing in embodied robots.
To compress robot development from months to weeks, Diguai introduces a three-layer toolchain: Cloud-based RoboGo handles data generation, training, quantization, and deployment; PC-based RDK Studio manages device connectivity and physical robot deployment; and board-level Moss handles task execution, with Moss AgentEngine coordinating across all three layers.
Notably, the Sunrise S600 has already gained recognition from over 20 leading customers and has begun mass production validation. Digua has also collaborated with more than 100 industry partners to achieve coordinated compatibility across components such as chips, modules, sensors, robot controllers, and data acquisition systems.
Magic Atom: Three models cover all scenarios, from stage performances to real-world deployments.
Magic Atom unveiled three strategic new products globally at WAIC: the flagship full-sized MagicBot X1, the industrial wheeled MagicBot D1, and the light industrial quadruped MagicDog T1. Also showcased were nine scenario-based solutions and the company’s proprietary general embodied large model, Magic-VLA K02.
The three products each serve distinct roles: the X1 stands 180 cm tall, features 31 active degrees of freedom, and delivers a peak torque of 450 N·m per joint, enabling it to reliably perform high-dynamic tasks such as dunking, table tennis, and fencing—pursuing the limits of general-purpose capability.
D1 has been deployed in the Mijia Intelligent Manufacturing Factory, featuring a wheeled chassis with a humanoid upper body, enabling autonomous navigation without tracks and cluster scheduling to create a fully automated loop for material handling, cross-area transportation, and loading/unloading on production lines. T1 complements this by addressing decentralized scenarios in light industry and campus environments.
Image source: Lei Technology production
The proprietary Magic-VLA K02 model demonstrated on-site execution of complex, long-duration tasks such as box sealing with adhesive, organizing flexible clothing, and fully packing a suitcase, achieving an overall integrated operation success rate exceeding 90%.
This model employs a two-layer collaborative logic: the upper layer first understands the complete task objective, while the lower layer seamlessly executes a series of precise actions, completely overcoming the limitations of early robots that could only rigidly replicate single steps, and enabling autonomous planning of the entire operational workflow with the final deliverable as the direct goal.
In terms of commercialization, Magic Atom is one of the most experienced players in the industry. Previously, its fleet of hundreds of robots appeared on the CCTV Spring Festival Gala and performed at international events. This April, it directly entered the public safety sector by partnering with the Wuxi Municipal Public Security Bureau, deploying its self-developed traffic management humanoid robot “Xiao Mai” to manage traffic at the Wuxi Marathon—marking one of the first official deployments of a humanoid robot in a government scenario for real-world operations.
Leju Robotics: Actually working in factories to move goods, with over 95% of components domestically sourced.
Leju Robotics set up a prominent "Brick-Moving Zone" at the exhibition, featuring a 1:1 replica of a real production line. Three humanoid robots tirelessly handled palletizing and depalletizing of cardboard and plastic boxes, as well as small-part loading. When boxes were misaligned, the robots automatically corrected their position—entirely without human intervention. These solutions have been deployed and operating in factories for three to four months, serving over ten leading enterprises across industries including FAW, ZTE, and Changhong.
Leju also showcased its end-to-end development toolchain, which significantly reduces deployment cycles: standardized exhibition scenarios can be deployed in 1 to 2 days, industrial scenarios in 1 to 2 weeks, and customized scenarios in 1 to 2 months. The logistics solution developed in collaboration with Ant Group’s Lingbo requires only 300 data samples and can be fine-tuned in 1 to 2 days, achieving a handling success rate of over 80%.
Image source: Lei Technology production
Notably, the localization rate of the Leju Kuafu series has exceeded 95%. By collaborating with Huixi Intelligence on a domestically developed high-performance main control solution, the final piece of the full-stack localization puzzle has been completed. Additionally, Leju’s “Intelligent Picking and Transporting Solution for Humanoid Robots in Automotive Manufacturing Scenarios” has been selected as a typical case by the Ministry of Industry and Information Technology for artificial intelligence applications.
Thousand Sensory Robotics: Adding tactile perception to distinguish authentic from counterfeit sneakers by touch
Qianjue is the most overlooked company at WAIC, yet what it does addresses the industry’s most critical短板. It aims to solve the more fundamental physical problem of robotic touch. Even after robots make contact with objects, they still see but cannot accurately feel or control them—lacking tactile feedback, they cannot cross the threshold into fine-grained operations.
Image source: Lei Technology production
During WAIC, Qianjue Robotics unveiled X-TouchMind V1, the first embodied haptic model designed for tactile intelligence, along with TacVerse 1k, a 1,000-hour visuohaptic multimodal dataset.
The booth features two demos side by side: one demonstrating the robotic formation of double-arm long sequence cartons, where the robot sequentially performs picking, unfolding, aligning, folding, and compacting on flexible cardboard; the other showcasing precision assembly of headphones in a 3C manufacturing scenario, where the robot uses haptic feedback to autonomously achieve precise alignment and dynamic correction during delicate operations.
At the Duwu app booth, this haptic intelligence system is already in operation. Equipped with Qianjue’s proprietary VTLA and haptic sensors, the robot completes fully automated shoe quality inspection in just five seconds, with identification accuracy exceeding 99.9999% compared to human experts.
This data cannot be obtained visually at all; Qianjue’s force sensing accuracy reaches the millinewton level, surface texture sensing reaches the micrometer level, and the gripping and pressing process is fully compliant and adaptive, causing zero damage to footwear.
Additionally, Qianjue showcased its proprietary XTac UMI G1 wearable haptic data acquisition gripper, which transforms real-world human operational data—such as contact variations, force feedback, and slip trends—into high-quality embodied datasets. The full-hand haptic solution establishes a comprehensive tactile feedback network from the fingertips to the palm using fingertip sensors and electronic skin modules.
Qiyuan Robot: Transformers come to your living room, centered on open-source DIY customization.
Under Weiwei New Materials, Qiyuan Robotics made its public debut at WAIC with the Qiyuan T1—the world’s first transformable personal robot. Featuring a Transformer-based unified architecture, it autonomously switches between wheeled humanoid and quadruped forms: acting as a humanoid assistant indoors and transforming into a robotic dog outdoors. It seamlessly integrates with motion cameras like Insta360 for voice-controlled filming, intelligent tracking, and smooth motion cinematography.
The Qiyuan Q1 Explorer Edition is now officially released, featuring a design that won the international A' Design Award Gold Prize. All structural components are fully open-sourced and support 3D printing for personalized customization.
Qiyuan doesn't just sell a robot—it's a hardware platform that users can customize with their own casings, add accessories, and modify functions, aligning more closely with consumer electronics logic.
Image source: Qiyuan
Modular design, open-source architecture, and 3D-printed customization are basic standards in the consumer electronics industry, bold experiments in robotics—but this new breed of product demands precisely such an adventurous spirit. The Qiyuan Q1 has now officially launched its market presence in North America and Europe, using design as a bridge to convey its value to users worldwide. Moving forward, the Qiyuan Q1 will continue pursuing its mission of “a robot for everyone,” becoming an integral part of more lives around the globe.
The robot brain takes center stage, with VLA and world models advancing in parallel.
No matter what novel models or flashy solutions each company introduces, the ultimate goal is always the same: to get robots working in the real world.
For a robot to successfully complete its tasks, it relies on two core components: first, an intelligent brain that organizes task logic and plans operational steps; second, a mechanical body that executes actions with precision.
The topic that industry professionals discussed the most and disagreed on the most at this year’s conference was which technological path the robotic intelligent brain should take.
Currently, the industry is primarily divided into two tracks: VLA (Vision-Language-Action) models and world models. Both have their respective limitations, and no unified framework has yet emerged that perfectly integrates the two.
The VLA model excels at understanding human natural language and quickly adapting to diverse scenarios, but its weaknesses are clear: it struggles to predict physical laws such as object deformation under force, friction, and motion. World models, on the other hand, can accurately simulate various physical interactions, but they are relatively weak in interpreting text and understanding complex human instructions.
Da Xiao Robot represents the pure world model approach, but its lead developer has publicly stated that current world model technology is still insufficient; for standardized, fixed tasks, VLA still learns faster and operates more stably, so both approaches will proceed in parallel in the short term.
He predicts that once robotic real-world operational data accumulates to a scale of tens of millions of hours, world models will gradually absorb all the capabilities of VLA, ultimately forming a unified foundational large model.
Image source: Lei Technology production
At the main forum of WAIC, Yao Maoqing, partner at Agi and founder of Mifeng Technology, also highlighted two major flaws in current world models:
First, there is a data type mismatch. The physical world data required by internet videos and robots are fundamentally different. Content on the internet may defy physical laws—people flying is perfectly acceptable—but robots need real-world physical interaction data such as pushing, pulling, twisting, gripping, flexible objects, friction, and fluid dynamics.
This type of data is scarce online, and existing physics engines struggle to simulate it.
Second, the path of world models is bottlenecked by data, not architecture. Large language models have already been trained on 100 trillion tokens, equivalent to 10 billion hours of speech. Data from the physical world has far lower information density than text; achieving advanced physical commonsense reasoning may require real-world data on the scale of billions of hours.
Yao Maoqing’s assessment is: scenarios with high-frequency necessity, controllable environments, strong certainty, and tolerance for error will be deployed first; more open environments and tasks, along with general capabilities, will follow in the next step.
Image source: Lei Technology production
Therefore, Zhiyuan Robotics has adopted a dual-track approach: on one hand, it relies on its self-developed GO-2 embodied foundation model as the decision-making core, combined with action chain-of-thought and an asynchronous dual-system architecture to ensure stable real-world deployment of the VLA model; on the other hand, it continuously iterates its GE-2 world model, which leads its category, to achieve an upgrade from virtual simulation to autonomous learning in the real world.
Simultaneously launching the Genie Evolver 1.0 real-world closed-loop reinforcement learning framework, pioneering a third differentiated approach: instead of relying on chaotic public internet video data for training, the robot continuously interacts in real operational environments, accumulates hands-on experience, and autonomously iterates and evolves.
When will world models mature? According to Wang Cong, CEO of Digu, it will be when people no longer consciously mention the term “world model”—this will indicate that the path has truly succeeded, just as today no one debates which technical processes autonomous driving uses; they only care about how well it actually performs.
Robots enter their "CATL moment," with the supply chain stepping into the spotlight.
This year, there was another very noticeable change at the exhibition: upstream supply chain companies—such as those producing chips, joints, and sensors—no longer played supporting roles behind major device brands; instead, they confidently stepped onto the stage to showcase their core products.
The pace of domestication is visibly accelerating: the Leju Kuafu series of humanoid robots have achieved a domestic component rate exceeding 95%. In collaboration with Huixi Intelligence, they have developed a domestically produced high-performance control solution, filling the final critical gap in full-stack domestication of humanoid robots. Their humanoid picking and transporting solution tailored for automotive factories has been selected as a typical case of AI implementation by the Ministry of Industry and Information Technology.
Image source: Lei Technology production
The Sweet Potato Robot Sunrise S600 compute module has collaborated with hundreds of upstream and downstream industry chain enterprises to achieve seamless hardware-software integration, covering all stages including chips, core modules, various sensors, robot motion controllers, and multimodal data acquisition.
In addition, Zhishen Technology has independently developed the CHAMP Champion series of joint modules, which are supplied not only to its own complete robot production lines but also to manufacturers across the industry. For years, joint components accounted for nearly 50% of the total cost of humanoid robots, with overseas companies monopolizing core technologies and production capacity; the domestic self-developed joint modules have directly broken through this critical bottleneck.
The overall domestication rate of core components for humanoid robots across all categories has now exceeded 75%, with a complete, self-owned domestic supply chain gradually taking shape—from underlying computing chips and actuated joints to multimodal perception sensors. This signifies that humanoid robots have fully transitioned from experimental prototypes assembled from scattered imported parts to mature industrial products capable of mass production and standardized delivery.
In addition to full systems, large models, and computing infrastructure, humanoid dexterous hands are also a fiercely competitive niche segment, with each company showcasing distinct approaches and perception solutions, all of which prominently unveiled their core hardware at this year’s WAIC.
Taking Lingxin Qiaoshou as an example, this exhibition features mature products from two mainstream technical pathways: the direct-drive Linker Hand O30 series, designed for complex, precise operations and compatible with training of various embodied manipulation large models; and the upgraded new tendon-driven dexterous hand, Linker Hand L30 Pro.
The two solutions correspond to different scraping and manipulation scenarios, covering requirements for scientific research and industrial flexible operations.
Image source: Lingxin Qiaoshou
The milestone also marked the domestic debut of the flagship OmniHand 3 Ultra-M, a fully direct-drive dexterous hand featuring 20 active degrees of freedom, with each fingertip equipped with an integrated visuo-tactile sensor and over 300 three-dimensional tactile sensing points covering the entire palm, enabling simultaneous capture of pressure, deformation, and sliding details of objects—essential hardware for high-precision contact-based operations.
In terms of underlying tactile infrastructure, Tashan Technology has established a complete tactile perception industry ecosystem, integrating self-developed tactile chips, sensing components, and end-to-end implementation solutions.
The flagship exhibit, the Emerald E10A chip, has been officially defined as the world’s first dynamic haptic sensing chip, featuring microsecond-level ultra-high time resolution and a standby power consumption of only 0.9μA. Leveraging the technological advantages of this chip, the company’s related haptic products have already captured over 80% of the market share in the humanoid robotics sector, establishing it as a core supplier of haptic hardware in the industry.
The robotics industry's supply chain is following the path taken by the electric vehicle industry—back then, CATL and BYD also went from being mocked to becoming global leaders. The "CATL moment" for robotics may not be far off.
No all-in-one players; the industry chain is accelerating specialization.
By now, readers will likely have noticed a trend: no company is fully focused on controlling every stage of the supply chain.
Two years ago, in the humanoid robotics sector, nearly all companies promoted their core strength as "end-to-end in-house development," aiming to independently complete everything from underlying chips and actuator drives to algorithm models and final assembly.
However, WAIC 2026 presents a completely new industrial landscape, where each company focuses on refining a specific niche rather than trying to dominate the entire value chain.
Chip makers should focus on chips; Digu Robotics does not manufacture complete devices—it is building the computing infrastructure and development toolchain. The Xuri S600 collaborates with over 100 industry partners, while UBTECH’s Cruzr Y1 provides the “brain.”
Let those focused on joints concentrate on joints—Champ Series joint modules from Zhishen Technology are not only used in-house but also sold externally. Joints once accounted for nearly 50% of the total cost of humanoid robots and were long dominated by foreign suppliers. Now, a company has emerged specifically to tackle this challenge, with annual production capacity planned at the million-unit scale.
Focus on haptics—Qianjue Robotics, founded less than two years ago, has secured three funding rounds to enable robots to “feel.” With its VTLA model, XTac UMI G1 data-capture gripper, and full-hand electronic skin solutions, the company does not build complete robots but specializes exclusively in haptics—the “final centimeter.”
Image source: Lei Technology production
Focus on implementing real-world scenarios: Agi, Magic Atom, and LooPeng are concentrating on deploying robots into factories, subways, police stations, warehouses, and production lines. They don’t need to build chips and joints from scratch—others in the supply chain have already done that.
Let operators focus on operations; Qintian Rent doesn’t sell robots, but once customers pay their monthly fee, the robots are ready to work.
The era when a single company handled everything—from chips to complete devices to operations—is over. This is not a bad thing; the smartphone industry is the best example: Qualcomm makes chips, Sony makes sensors, Apple makes the operating system, and Foxconn handles assembly. No one owns the entire chain, but each company excels in its own segment.
The robotics industry is on this path—specialization is not a step backward, but a sign of maturity. At WAIC 2026, every company found its place, and that’s the industry’s biggest good news.
In conclusion
After exploring the entire 2026 WAIC, the biggest takeaway was that this year, everyone’s focus has shifted squarely to practical industry issues—such as production capacity, delivery, implementation costs, and scenario replication rates. Humanoid robots have left the lab and are now entering our production systems, public services, and even daily life.
Of course, the industry still faces several unresolved challenges: limited adaptability to complex home environments, potential for further cost reductions in certain core components, and the absence of a unified integrated model. However, it is undeniable that the industry has now completed its most difficult phase of initial exploration from zero to one and has entered a window of scalable expansion from one to a hundred.
Companies are no longer working in isolation or engaging in blind internal competition; instead, they are leveraging a complete domestic industrial chain to collaborate and complement each other—some focus on foundational components, others on hardware, some on integrated systems, and others on operations, each playing their part and advancing steadily.
In 2026, this conference serves as a turning point: humanoid robots have entered the workforce and become productive tools continuously generating real value, marking the true beginning of the golden era of large-scale commercial deployment of humanoid robots.