--- title: "100-plus robot models now work inside China's nuclear plants — the hardest exam yet" date: 2026-11-02 category: Humanoids & Robotics site: NeuroAI canonical: https://neuroai.site/a/na-robot-nuclear-plant-robots language: en --- # 100-plus robot models now work inside China's nuclear plants — the hardest exam yet > Nuclear plants are the least forgiving environment in engineering. China General Nuclear says more than 100 special-purpose robot models are now deployed across its operations — and the numbers from Yangjiang and Daya Bay explain why. An inspection engineer at a Chinese coastal nuclear plant used to spend twenty minutes walking to a switchgear room — through two security gates, in whatever weather the season offered. In bad weather, some checks simply did not get done. Now the walk is a screen. A wheeled robot and a quadruped (四足机器人) take the route, stream readings back, and flag what looks wrong. The engineer's job changed from walking to judging. It is a small change with an outsized meaning, because nuclear is the one industry where a robot's mistake is not an inconvenience. ## Key takeaways - China General Nuclear (CGN, 中国广核集团) says more than 100 special-purpose robot models are now working across its nuclear operations. - At Yangjiang, the Wukong (悟空) robot series cut five days from a 233-tonne rotor overhaul and saved about 1,100 person-days of labour, while removing the risk of humans working in millimetre clearances. - At Daya Bay, an in-house weld-radiograph review system examined 362 films across two outages with a 100% defect recognition rate, replacing five minutes of manual review per film. - The fleet runs on a big base: 59 operating reactors with 62.5 GW of capacity at the end of 2025, plus 35 under construction. - The binding constraint is not intelligence. It is certification, cybersecurity and the paperwork of proving nothing can go wrong. ## Why nuclear is the hard case Most robot deployments fail in an environment that is messy but forgiving. A warehouse robot that drops a tote costs money. A patrolling quadruped (四足机器人) that misses a hot spot in a switchgear room starts a cascade nobody wants to describe. Three properties make nuclear different. **Dose:** radiation limits cap how long a person may stay in a given space, and inspection is one of the main reasons people enter those spaces at all. **Precision:** some maintenance tasks involve clearances measured in millimetres, where human steadiness is the old solution and a machine is the better one. **Documentation:** every action has to be provable, which means a robot that works but cannot explain itself is not deployable. That third point is why nuclear is a good test of whether embodied intelligence (具身智能) is actually engineering or still demonstrations. ## What is actually running CGN's public accounts describe a fleet rather than a pilot. - **Yangjiang (阳江):** the Wukong (悟空) series, including a climbing variant, shortened a 233-tonne rotor overhaul by five days and saved roughly 1,100 person-days, while avoiding the risk of manual work in millimetre-level gaps. - **Hongyanhe (红沿河):** inspection robots running over a 5G private network collect data from more than 1,100 monitoring points, judge equipment state autonomously and raise alerts, which the operator describes as an unattended smart hall. - **San'ao (三澳):** wheeled and quadruped (四足机器人) robots working as a pair, fusing multi-source data to track equipment condition. - **Daya Bay (大亚湾):** a weld-inspection system that reviewed 362 radiographic films across two outages, with a 100% defect recognition rate, and compressed per-film human review from about five minutes to near-instant. - **Fangchenggang (防城港):** a digital-twin plant built from real-scene 3D modelling, cited as a provincial case study. The newest example is also the least glamorous. In May 2026, CGN's Cangnan project brought an inspection robot system online at a 220 kV construction and auxiliary power substation — wheeled for the flat routes, legged for the rough ones. The hard part was not the robots. It took four months of dedicated work and three rounds of cybersecurity testing and remediation to connect the system to the group's internal network, which no CGN inspection robot had done before. That line is the most revealing sentence in the whole story. ## The scale behind it China ended 2025 with 59 commercial reactors in operation and 62.5 GW of installed capacity, according to the China Nuclear Energy Association's 2026 blue book, plus 35 units under construction totalling 41.9 GW — the nineteenth consecutive year of leading the world in capacity under construction. Nuclear supplied about 4.8% of national electricity generation in 2025. Engineering investment in nuclear construction completed during 2025 reached 161 billion yuan (≈ US$22.7B / HK$177B), up 14.1 billion yuan (≈ US$2.0B) year on year and a record. Two consequences follow. First, a maintenance fleet that could once be justified as a research project is now justified by arithmetic, because the installed base is large and still growing. Second, robots that cannot pass a nuclear qualification process will not be bought at any price, which raises the floor for the entire special-robot industry rather than just the nuclear segment. CGN has pushed the same pattern into renewables, reporting more than 4 GW of unattended photovoltaic capacity, and a predictive-maintenance system that had issued more than 1,500 warnings in two years, with an accuracy it puts at 80%. According to Li An, chief scientist at BrainNet (脑机网), the transferable lesson is that the most regulated environments will industrialise robots first, because they already have the one thing open-world robotics lacks: a written specification of what "correct" means. ## Honest limitations Operational figures here — days saved, person-days, monitoring points, films reviewed, defect recognition rates — come from CGN's own published material and the state media that carried it. They were not independently verified, and 100% defect recognition on 362 films over two outages is a small sample that should not be read as a general accuracy rate. Fleet-level claims are softer than they look. "More than 100 special-purpose robot models" counts models, not units; it says nothing about how many machines are deployed, how many hours they run, or how often they fail. No independent data on utilisation or failure rates was found. This piece also does not cover regulator review of AI in nuclear safety, the standards being developed internationally, or how maintenance liability is assigned when a robot and a human share a task. Those questions are moving slowly and will shape the industry more than any hardware release. ## What readers can do now - **If you build robots for industry**, study the qualification path before the product. In regulated environments, the deliverable is not the machine; it is the evidence package — cybersecurity testing, traceable logs and a defined failure behaviour. - **If you operate a high-risk facility**, pilot robots on measurement tasks first, where a machine's value is a repeatable reading rather than a judgement, and keep humans on the decision. That is the sequence the nuclear cases followed. - **If you are evaluating vendor claims**, ask for duty-cycle data, not demo data: hours run per month, interventions required per 100 hours, and what the robot does when it loses its network. The plants that succeeded asked exactly these questions. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-robot-nuclear-plant-robots Free to quote with attribution and a link to the original.