China Southern Power Grid, the state-owned utility supplying electricity to 272 million people across five Chinese provinces, is training humanoid robots to perform substation inspections and basic electrical maintenance. At a laboratory within the company’s Guangdong operation, a cohort of robots is learning by mimicking young engineers – opening and closing control boxes, switching equipment on and off, and checking for electrical leakage using insulating rods.

The program reflects a broader effort by Chinese state enterprises to deploy robotics in infrastructure roles where labor demand is high, conditions are physically demanding, and error carries significant operational consequence.

What the Robots Are Being Trained to Do

The current training program covers six operational scenarios and 15 discrete skills. Robots have already been deployed and field-tested in some of these scenarios since last year. Tasks include inspecting transmission line insulators, detecting fractures in steel cores, and tightening screws on elevated equipment – work that must be carried out regardless of weather conditions.

“Taking a 500 kV substation as an example, there are thousands of pieces of equipment to inspect and nearly ten thousand inspection points,” said Li Duanjiao, head of the company’s robotics laboratory. “The work must be carried out under intense sunlight, in the rain, or during storms. When performed manually, the workload is enormous and highly repetitive.”

The utility uses a mixed fleet. Humanoid robots handle manipulation tasks at control equipment. A quadruped robot named Feiyun, capable of carrying loads up to 100 kilograms, transports robotic arms into field locations – a logistics role previously performed by human workers. Drones carry out aerial inspections. The company refers to these systems collectively as “digital employees”.

The Infrastructure Case for Robotics

Power grid maintenance is a natural fit for the current generation of industrial robots. The work is structured enough to be codified into repeatable sequences, physically demanding enough to strain human workforces, and geographically distributed enough – particularly in remote mountainous terrain – that labor deployment is expensive and slow.

Li said robots are gradually replacing humans in the most repetitive and remote inspection tasks, with the program expanding as training data accumulates from field deployments.

UBTECH Robotics, a Shenzhen-based private company, has taken a parallel approach in automotive manufacturing, deploying robots to transport materials and select components on car production lines. The company also developed one of the first robots capable of autonomous battery swapping. Tan Min, a company representative, estimated that five to ten years remain before robots are widely adopted across industries at scale – a timeline he tied directly to the volume of real-world training data still needed to advance the underlying AI.

Data as the Binding Constraint

The power grid program illustrates a challenge that applies across the sector: robots improve through real-world exposure, but real-world deployment requires robots capable enough to operate safely. China Southern Power Grid’s laboratory approach – structured mimicry training followed by controlled field deployment – is one model for managing that transition.

With more than 150 humanoid robot manufacturers active in China, the near-term competition is not primarily over hardware specifications but over which companies and institutions can accumulate sufficient operational data to make their systems reliable in unstructured, high-stakes environments. Infrastructure operators like China Southern Power Grid, with clearly defined task sets and existing engineering workforces to train from, are positioned to generate that data faster than most.

X Square Robot, a Shenzhen-based embodied intelligence company, has launched a pilot home-cleaning service in partnership with household platform 58.com, deploying humanoid robots alongside professional cleaners in residential settings across the city. The robots handle structured, repeatable tasks – wiping surfaces, collecting debris, tying garbage bags, and assisting with bedsheet folding – while human cleaners manage work requiring judgment and adaptability.

The pilot has drawn attention from the international robotics community as an early commercial test of embodied AI in one of the most operationally complex environments available: private homes.

Why the Home Environment Matters

Home cleaning represents a deliberate choice as a deployment context. Unlike factory floors or logistics warehouses, residential spaces are unstructured, variable, and unpredictable – no two homes are configured the same way, and conditions change between visits.

“The service industry has extremely high complexity and non-standard characteristics, and the home environment is regarded as the ultimate benchmark for evaluating general robots,” said Yang Qian, chief operating officer of X Square Robot. The company uses an end-to-end AI model that allows robots to interpret tasks, plan multi-step actions, and execute them autonomously – without scripts or remote control.

The hybrid deployment model – one robot, one human cleaner per booking – is also a deliberate design choice. Chris Paxton, an AI and robotics researcher, described the approach as one that “allows you to scale from 70 percent to 90 percent to 99 percent autonomy naturally”, as robots accumulate real-world data and improve over time.

Shenzhen as a Robotics Testing Ground

The pilot is operating within a broader industrial ecosystem. Shenzhen has gathered over 2,600 AI enterprises and more than 70,000 companies across the robotics supply chain. In 2025, the combined added value of Shenzhen’s AI and robotics clusters grew at a double-digit rate.

The city’s Robot Valley, where X Square Robot is headquartered, houses companies spanning hardware manufacturing, sensor development, and AI software. Pudu Robotics, another Shenzhen-based firm, now serves clients in more than 80 countries, with overseas markets accounting for over 80% of its total revenue for several consecutive years. Its commercial cleaning robots have found particular traction across Europe, North America, and Asia.

Beyond commercial applications, robots have begun appearing in Shenzhen’s public infrastructure. Robot volunteers have been deployed at Qianhaishi Park, and autonomous units have been observed directing traffic – deployments that reflect the city’s role as a live testing environment for a wide range of robotic applications.

Data and the Path to Scale

For X Square Robot, the 58.com pilot serves a dual purpose: generating revenue while accumulating the real-world operational data needed to train and refine its AI systems. Each cleaning session produces feedback on task execution, environmental variation, and edge cases that controlled environments cannot replicate.

The broader direction of China’s robotics sector is toward domestic service and eldercare alongside industrial applications – markets that require robots capable of operating reliably in unstructured, human-occupied spaces. The Shenzhen pilot is an early data point in that effort, and its results will likely inform how quickly similar deployments expand to other cities and service categories.

Agibot has deployed its G2 humanoid robots in an active production environment at Longcheer Technology, a Shanghai-based consumer electronics manufacturer. The rollout marks one of the more concrete demonstrations of humanoid robots operating within a live industrial workflow, moving the technology beyond controlled pilots into continuous factory-floor use.

The deployment comes months after Agibot announced the production of its 10,000th humanoid robot in March, a milestone the company described as a turning point in the industrialization of embodied AI.

What the Robots Are Doing

G2 units are stationed at multimedia-integrated testing stations, where they perform precision loading and unloading of devices into testing fixtures. The task demands millimeter-level placement accuracy, consistent cycle timing, and the ability to sort finished and defective units without interruption.

Agibot reports throughput of up to 310 units per hour, with individual cycle times of approximately 19 to 20 seconds per operation and a success rate exceeding 99% in continuous use. Production line integration was completed within 36 hours, and each shift produces approximately 3,000 units. The system has logged over 140 hours of cumulative continuous operation, with downtime losses below 4%.

The robots use multi-modal sensing – combining visual perception and spatial awareness – to identify objects and execute task sequences without custom tooling. The platform supports mixed-model production, meaning it can handle different device configurations on the same line, reducing changeover time.

The Underlying Architecture

Agibot describes its approach as a full-stack ecosystem for embodied intelligence, integrating robot hardware, AI models, and large-scale data infrastructure into a single system designed for continuous learning. Rather than executing fixed instructions, the robots are built to adapt to task and environment variations over time through a combination of simulation-based validation, reinforcement learning, and on-device inference.

“This project shows that embodied AI is no longer experimental. It is a practical, production-ready capability that can operate reliably in real industrial environments and deliver measurable economic value,” said Maoqing Yao, Partner, Senior Vice President, and President of the Embodied Business Unit at Agibot.

Scale and Next Steps

Agibot plans to expand the deployment to 100 robots by Q3 2026 and has identified automotive, semiconductors, and energy as the next target sectors. The company is also developing Genie Sim 3.0, a simulation platform designed to accelerate training of new robot behaviors before physical deployment.

The broader implication of the Longcheer deployment is not the throughput figures alone, but what they suggest about deployment speed. A 36-hour integration timeline and no custom tooling requirement lower the barrier for manufacturers evaluating humanoid robots against conventional fixed automation. Whether those numbers hold across more varied factory environments – with different layouts, device types, and production rhythms – will determine how transferable the model is at scale.