Executive Summary
At the 2026 World AI Conference (WAIC) in Shanghai, "embodied intelligence" remained the dominant buzzword, but away from the demo stages, engineers and operators disclosed harder numbers. Singapore's Tuas Port has more than 530 magnetic-guided AGVs in operation. China's ports now run more than 800 autonomous intelligent vehicles (AIVs). Changzhou First People's Hospital has put a domestically built bronchoscopy navigation robot and puncture robot into clinical use. A robotics company serving Apple's supply chain has delivered hundreds of six-arm industrial robots for inspection tasks. At the same time, several operators acknowledged how far the industry still has to go — China Logistics Group stated that no company has yet achieved reusable, cross-embodiment robot data. This article draws on transcripts from four WAIC 2026 forums covering ports, logistics, healthcare and industry to document the actual scale, performance figures, and unresolved engineering problems behind current robot deployments.
Industry Context
WAIC 2026's exhibitor mix shows robotics has moved from a technology showcase into a supply-chain category of its own. Wang Bo, Chief Content Officer at Jiazi Guangnian, disclosed at the Shanghai MaQiao AI Innovation Zone "AI Transformation" Embodied Intelligence Industry Ecosystem Forum (the "MaQiao Forum") that 242 robotics, embodied-intelligence and smart-hardware exhibitors attended this year's WAIC — 23.6% of all exhibitors.
Capital inflows point the same direction. Zhang Yijia, Founder and CEO of Jiazi Guangnian, released the "2026 China Embodied Intelligence Industry Insight" report at the same forum, citing 322 financing deals in embodied intelligence in the first half of 2026, totaling more than RMB 90 billion, with roughly 20 new unicorns added in that period. China's dexterous-hand sales grew from about 19,000 units in 2025 to a projected 70,000 units in 2026.
Evidence of scaled deployment, however, remains uneven. Wang Tianmiao, Honorary Director of the Robotics Institute at Beihang University, noted at the MaQiao Forum that roughly 300 humanoid robot models appeared at this year's WAIC, while an estimated 80% gap remains in usable training data across the industry. Shao Tianlan, Founder and CEO of Mech-Mind Robotics, offered a comparison at the WAIC 2026 Intelligent Trends Forum, "Toward Autonomy: The Real Challenges and Solutions of Industrial AI Deployment" (the "Industrial AI Forum"): roughly 1 billion people worldwide work in manufacturing and logistics jobs, while annual shipments of industrial plus collaborative robots total only about 500,000 units.
Technology: From Single-Unit Intelligence to "Brain Plus Cerebellum"
Several forums independently used a "brain plus cerebellum" framing to describe the current technical bottleneck. Cheng Liang, General Manager of the Central SOE Business unit at Amap, said at the "AI-Enabled High-Quality Development of the Logistics Network" Forum (the "Logistics Forum") that most embodied-intelligence robots on display still rely on human teleoperation to perform demonstration actions such as dancing or fine manipulation, and lack the ability to move autonomously in the real physical world — what he called a missing "cerebellum." Amap used this WAIC to launch its general-purpose world model workshop, "Able World Studio," and has piloted it on the company's self-developed quadruped robot, targeting guide-dog services for China's 17 million visually impaired citizens — an application not yet formally launched.
Qian Feng, Academician of the Chinese Academy of Engineering and professor at East China University of Science and Technology, proposed a framework of "industrial embodied intelligence" at the Industrial AI Forum: motion programs for factory PLCs, DCS systems and robots are still entirely written by humans today, and equipment, workshop sections, production lines and entire factories need an "industrial brain" to move from automation based on preset rules toward autonomy based on self-learning optimization. He cited a 40-million-ton-scale oil refinery where an AI "brain" now closes the loop across planning, real-time scheduling and unit-level control.
In the port sector, Zhai Shaopeng, a young scientist at Shanghai AI Laboratory, described a "port world model" at Shanghai International Port Group's "AI-Enabled Smart Port Global Shipping Ecosystem" Forum (the "Port Forum"): borrowing the large-language-model paradigm, the model treats terminal operating system (TOS) instruction sequences as tokens and uses a Transformer to autoregressively predict the next TOS instruction. Trained on a year of Yangshan Phase IV data — tens of millions of records — the roughly 2-billion-parameter model improved overall performance by about 5% when combined with existing smaller models. Zhai said candidly that the port world model is still at its "GPT-1 moment," with substantial room to grow in model scale, dataset size and application paradigms.
Engineering Analysis: Reliability Before Form Factor
Speakers at the Industrial AI Forum repeatedly stressed that factory environments tolerate almost no error. Liu Zhen, President of the Digital Building Materials Research Institute (under China National Building Materials Group) and an IEEE Fellow, said his division's automatic run rate rose from 10%–20% at plant entry — when customers would cut over to manual control at any sign of deviation — to more than 95%, with a target of 100%; an additional "model of the model" layer was built on top for safety governance.
Xu Ximing, Senior Vice President at Hikvision, said the company's trinocular structured-light camera for robotic arms passed safety certification with a 99.6% unboxing accuracy rate and has entered strategic partnerships with Midea and KUKA; perception false-alarm accuracy rose from 93%–94% to 99%. He also pointed to an economic constraint: summarizing one video stream at one frame every five seconds using a general-purpose large model costs about RMB 4,000 per month in inference; Hikvision compressed that to about RMB 20. Without that reduction, he said, replacing a single "carbon-based" worker with 50 video channels at RMB 200,000 a month in inference cost would not make financial sense for any company.
Reliability claims in physical AI still lack a common evaluation standard. Cao Yunan, Chairman and CEO of Elite Robotics, said at the Industrial AI Forum that physical AI, mapped against autonomous-driving maturity levels, sits at roughly L2 industry-wide; he noted that companies routinely claim 99% success rates in specific factories or scenarios without any independent third-party verification. Shao Tianlan's product-maturity model at Mech-Mind places most public demos at levels 0–3 on a 0–9 scale, with level 8 marking a genuine revenue-generating business and level 9 marking profitability; he assessed that large-scale deployment in manufacturing and logistics is accelerating, large commercial services are approaching that threshold, and home use remains further out.
Engineering challenges at ports are similarly concrete. Chen Lianghui, Vice President of Tuas Port at Singapore's PSA International, said mechanical removal and fastening of container "lashings" — widely regarded in the industry as a frontier challenge in terminal automation — is currently about 60%–70% mechanized, with AI being applied to optimize the remainder. Heng Liang, Chief Scientist at Youdao Zhitu, said the marginal gains from improving any single vehicle's onboard intelligence are now small; the key going forward is building a cloud-based perception-fusion system that unifies data across different vendors' vehicles and human-driven vehicles into a single traffic picture — a shift from "single-vehicle intelligence" to "swarm intelligence."
Commercial Progress: Who Has Actually Deployed What, and at What Scale
Ports. Chen Lianghui disclosed that PSA's Tuas Port Phase 1 already operates more than 530 magnetic-guided AGVs, one of the largest such fleets in the world; Phase 2 will add roughly 900 autonomous-navigation IGVs running alongside the existing AGVs, with a combined annual throughput target of 65 million TEU once fully built. Youdao Zhitu's Heng Liang said China's ports now operate more than 800 AIVs (over 900 globally), a roughly 5% penetration rate, though newly built container terminals now commonly adopt autonomous vehicles as their primary horizontal transport. The company's own fleet of 250-plus AIVs runs in regular operation across ports in Shanghai, Quanzhou, Shenzhen, Shandong and Peru, and its smart trucks on the East Sea Bridge received four public-road driverless operating permits in July 2025. Shanghai Port Group handled 55.06 million TEU in 2025, ranking first globally for a 16th consecutive year; Yangshan Port Phase IV alone handled 8 million TEU in 2025.
Logistics. Zhang Yingying, Deputy Director of the Science and Technology Innovation Department at China Logistics Group, said the company's "Liuyun" AI model is already applied to AI-based processing of automated loading/unloading robots in warehouse scenarios such as wine storage, but acknowledged that robots remain "a long way" from actually working the factory floor, and that no company in the industry has yet achieved trainable, reusable data across different robot embodiments.
Healthcare. Zhou Jun, President of Changzhou First People's Hospital, described the robot fleet already in clinical use: outpatient companion robots; China's first domestically made bronchoscopy navigation robot (deployed in late 2024); China's first puncture robot (launched December 2025, providing real-time monitoring of puncture angle and depth); a da Vinci surgical robot; a ward-round robot that follows physicians and handles imaging retrieval, progress notes and order entry; a logistics robot; and drone-plus-robot-dog security patrols. The hospital's medical drone logistics network, launched with its first route in August 2024, now runs four routes on a regular basis. Cai Wei, President of Xinhua Hospital (affiliated with Shanghai Jiao Tong University School of Medicine) and the forum's moderator, added that Ruijin Hospital's drone delivery of frozen specimens cut delivery time from 20 minutes by courier to 2 minutes.
Industrial manufacturing. Zhang Jianzheng, Chairman and CEO of Saizhi Intelligent, disclosed that the company has become a direct supplier to "Company A" (Apple), with hundreds of six-arm industrial robots deployed in the lid-closing inspection step of smartphone final assembly — a task that employs roughly 350,000 workers globally; compressing that workforce to 100,000 would imply potential demand for roughly 200,000 robots. In a cabinet-inspection case for an NVIDIA IDC, the company said a task that took a worker 2.5 hours at a 90% success rate now takes a robot 6 minutes at 99.9% success — a case included in last year's World Economic Forum Global Lighthouse Factory whitepaper as its only embodied-intelligence application. Mech-Mind's Shao Tianlan disclosed that the company's "eye-brain-hand" embodied solution has cumulatively deployed more than 27,000 units, serving over 100 Fortune Global 500 customers, holding the top China market share for six consecutive years and leading markets in Japan, the US, South Korea and Europe, with compatibility across more than 40 robot brands and 1,000-plus models. Liu Jinming, Product Lead at JAKA Robotics, said the company has delivered more than 10,000 collaborative robots across 3C, automotive and parts industries since entering the space in 2015; this WAIC it launched a compact humanoid platform called TaiZai (1.22 meters tall, 27 degrees of freedom). Sheng Wei, Partner and Head of Product and Ecosystem at Tyrone Robotics, said the company's annual joint-module capacity is about 200,000 units and it already supplies harmonic-drive joints in bulk to most of the leading humanoid robot makers on the show floor.
Liu Zhen of the Digital Building Materials Research Institute disclosed that its high-quality industrial dataset exceeds 500 million records, with more than 2 million new records collected daily across 300-plus data sources; per-plant economic benefit ranges from several million to tens of millions of RMB, with payback periods of months or even weeks. A roundtable at the Industrial AI Forum disclosed that China National Building Materials Group has achieved real-time closed-loop AI control across single or full production processes in more than 100 factories over the past three years, and that a typical industrial enterprise with roughly RMB 10 billion in output value spends about RMB 80 million on first-phase AI infrastructure alone. Li Mingyang, Chairman and CEO of JAKA Robotics, said at a panel that customers' ROI expectations for embodied robots track those for six-axis and collaborative arms — generally a payback period within 24 months, with the most aggressive customers demanding payback in 10 to 12 months. Yu Kai, Founder and CEO of Horizon Robotics, said at the AI Empowers Ocean Industry Forum that its subsidiary D-Robotics is one of two major embodied-intelligence chip solution providers in China (the other being NVIDIA) and now supplies compute chips to more than 20 domestic humanoid and embodied robot companies.
Market Perspective
Multiple industry figures emphasized return on investment over technical showmanship when assessing the pace of commercialization. Li Mingyang of JAKA Robotics said at the MaQiao Forum's closing panel: "Your question already contains the answer: return on investment — that is forever the core." He added that customers are often more aggressive than robotics vendors themselves, sometimes requesting equipment reconfiguration after every 500 units of a product run.
Zhang Yijia of Jiazi Guangnian summarized the shift in competitive focus at the same forum: "Industry competition is shifting from who can build robots to who can stably work in real scenarios, who can continuously collect data, and form a commercial closed loop."
At the Industrial AI Forum, Liu Likang, Senior Vice President of Digital Industries at Siemens China, laid out a more conservative deployment path: build the IT/OT data foundation first, validate ROI at a single point, and only then pursue replication at scale, stressing that industrial AI should be "small but effective" rather than a grand narrative.
Challenges
Cross-embodiment data reuse remains a widely acknowledged weak point. Zhang Yingying of China Logistics Group stated plainly that no company has yet achieved trainable, reusable data across different embodied-intelligence platforms, and called for collaboration with technology companies to close that gap.
Multi-robot coordination at ports is similarly early-stage. Youdao Zhitu's Heng Liang said the key determinant of overall fleet efficiency is no longer improving any single vehicle's intelligence, but coordinating multiple vehicles — which requires a cloud-based perception-fusion system spanning different vendors and vehicle types. Xie Haiqin, Deputy General Manager at Caos IoT, raised a parallel point at the Industrial AI Forum: "Nowadays, we have more and more robots, more robot forms, and more brands. Although there is some industry consensus, I think it's very difficult to truly achieve real scheduling."
Hospital deployments carry particularly direct safety requirements. Zhou Jun of Changzhou First People's Hospital stressed that embodied robots used in hospitals "cannot keep making mistakes" — a precondition for real deployment in clinical roles — and said the hospital has adopted three principles: local data processing for security, no direct use of foundation models for patient diagnosis or treatment, and obtaining patient and ethics-committee consent.
Immature supply chains also constrain scaling. Sheng Wei of Tyrone Robotics said coordination among component suppliers remains difficult industry-wide, prompting the company to develop its own motors, reducers, encoders, drive boards and cross-roller bearings in-house for an integrated joint design.
RobotToday Analysis
Further reading: The port, factory and hospital deployments documented here correspond to the Field AI/Factory AI application branches and the dual-system "brain plus cerebellum" architecture mapped out in RobotToday's companion framework piece, Physical AI Landscape: From Digital Intelligence to the Embodied Physical World.
The figures disclosed at WAIC 2026 suggest robots have achieved genuine, sustained-scale deployment mainly in settings with well-defined task boundaries and quantifiable ROI: horizontal transport at ports, specific inspection and assembly steps in manufacturing, and standardized support tasks in hospitals. These settings share clear task scope, bounded failure cost, and mature data-collection pathways — which helps explain how Saizhi Intelligent moved an Apple supply-chain inspection task to hundreds of deployed units, and how PSA Singapore scaled its AGV fleet past 500 units.
By contrast, tasks requiring cross-scenario generalization, multi-robot coordination or long-tail manipulation — warehouse logistics or home service robots — remain earlier-stage. China Logistics Group's statement that no company has achieved reusable cross-embodiment data, paired with Mech-Mind's comparison of 1 billion manufacturing and logistics workers against 500,000 annual robot shipments, points to the same conclusion: the gap between demonstrations and scaled commercial deployment currently sits mainly in data reusability and engineering reliability, not in algorithms themselves.
Notably, several leading companies are describing their own maturity with more caution. Elite Robotics acknowledged its physical AI is at "L2, not even L3," while Mech-Mind's maturity model places most public demos at levels 0–3 out of 9. That language is itself a signal: the industry is developing a more explicit tiering vocabulary to separate systems that can demonstrate from systems that can generate sustained revenue. For buyers, the operating hours, failure rates and payback periods that ports, hospitals and factories are now disclosing may prove more useful benchmarks than any stage demonstration.
Quotes and figures in this article are drawn from conference public remarks at WAIC 2026 and have not been individually verified with the speakers or their organizations. Where a speaker's name could not be confirmed from the source transcript, this is noted in the text. Corrections are welcome.
WAIC Review 2/4: World Models vs. VLA: Robotics' Next 'ChatGPT Moment'
WAIC Review 4/4: From Port AGVs to Hospital Night Shifts: What WAIC 2026 Reveals About Real Robot Deployments
Leave a comment