Ruoyu Technology Takes Jiutian Into Energy Operations, but Safety Evidence Now Matters More Than AI Ambition

Ruoyu Technology presented its Jiutian robot brain to major energy companies on August 12, moving its embodied AI pitch into one of robotics’ least forgiving markets.

The Shenzhen startup attended the 2026 Energy Industry Technology Innovation and Integration Development Conference in Beijing. The event brought together representatives from PetroChina, Sinopec, CNOOC, China Huaneng, Huawei, China Mobile, and other organizations. Ruoyu founder and CEO Sun Teng also delivered a presentation about using embodied intelligence in hazardous industrial environments.

The appearance matters because Ruoyu is no longer presenting Jiutian as a general research platform. It is positioning the system as an intelligence layer for robots that inspect equipment, manipulate fuel hardware, and eventually respond to abnormal conditions. Those jobs expose every perception error, planning failure, and integration weakness to physical consequences.

The central contest is therefore not Ruoyu against one larger robotics company. It is adaptive, model-driven autonomy against the deterministic automation already trusted inside energy facilities. Jiutian promises greater flexibility, but industrial buyers will demand evidence that flexibility does not weaken safety.

Ruoyu Put Its Robot Brain in Front of the Industry That Can Test It Hardest

The Beijing presentation moved Jiutian closer to operational scrutiny, even though it did not establish a commercial deployment with every company in attendance.

According to the initial conference account, Ruoyu brought Jiutian and an embodied intelligence system for hazardous energy environments to the August 12 gathering. Sun explained how the robot brain supplies intelligence to machines intended for specialized operating conditions.

The distinction between attendance and partnership is important. Huawei, PetroChina, Sinopec, CNOOC, China Huaneng, and China Mobile participated in the broader conference. Publicly available information does not show that every listed organization adopted Jiutian or entered a commercial agreement with Ruoyu.

The conference itself ran from August 11 through August 13. Its published agenda covered oil and gas operations, electricity, renewable energy, equipment development, and industrial digitalization. The event program also identified major state energy groups and infrastructure operators among the intended participants.

That audience gives Ruoyu access to organizations with real facilities, operational data, integration teams, and procurement budgets. It also creates a much higher standard than a controlled demonstration.

A warehouse robot can stop when it encounters an unexpected box. A robot near fuel equipment must recognize the object, understand the operating state, respect exclusion rules, and place its hardware safely. A plausible action is not necessarily an acceptable action.

Ruoyu describes Jiutian as a multimodal robot brain. Multimodal means the system combines several input types, including video, audio, text, point clouds, and force feedback. The company says the software uses those inputs to understand scenes, interpret instructions, plan tasks, and produce motion trajectories.

That structure aims to connect three layers that conventional automation often separates. Perception identifies the environment. Planning selects a sequence of actions. Execution translates that sequence into robot movement.

Jiutian is also presented as hardware-flexible. Ruoyu’s public materials describe a “one brain, multiple bodies” approach that can connect the software to different robot forms. The company offers an edge interface for further development and lists an onboard computing configuration rated at 100 TOPS.

TOPS measures trillions of operations per second, but it does not measure useful robot performance. Processing capacity only indicates how much computation the device can theoretically perform. It does not reveal task success, response latency, failure frequency, or safe recovery behavior.

Ruoyu has not published enough independent test data to answer those questions across energy operations. Its Beijing appearance should therefore be read as an industrial positioning event, not as proof of broad field readiness.

What changed is still meaningful. The company placed Jiutian before the organizations that control many of China’s most demanding industrial environments. Those potential buyers can test whether a general robot brain survives contact with narrow passages, reflective equipment, weak connectivity, hazardous materials, and strict operating procedures.

The next challenge begins when a presentation becomes a procurement requirement. Energy companies will ask for defined tasks, measurable acceptance criteria, and a clear division of responsibility between model software, robot hardware, and facility controls.

Energy Operators Are Already Deploying Narrower Robots

Ruoyu is entering an energy market that already values robotics, but current deployments favor constrained missions with clearly defined safety boundaries.

The industry is not waiting for general-purpose embodied AI. Energy operators already use specialized machines for inspection, sensing, monitoring, and data collection.

In May, PetroChina reported that an explosion-protected quadruped had completed an intelligent inspection deployment at the Xigu oil depot in Gansu. The robot inspected tank bases, pump areas, slopes, and trenches while sending data into an operational platform.

That oil-depot deployment used multiple positioning stations to address signal obstruction and drift around metal tanks. Engineers also adjusted the robot’s circuitry, enclosure sealing, and navigation behavior for the site.

This example establishes a practical benchmark for Jiutian. The quadruped did not need to perform every human task. It needed to traverse a known facility, gather defined information, and complete repeatable inspections without introducing new hazards.

A separate PetroChina operation placed an intelligent inspection robot into regular use at the Alashankou metering station on the China-Kazakhstan oil pipeline. The machine carries an explosion-protected camera, infrared imaging, and combustible-gas detection equipment.

The robot conducts automated patrols, collects readings, checks for leaks, creates reports, and sends warnings to a back-end system. Again, its mission is narrow enough for operators to specify expected behavior.

These deployments pressure Ruoyu in two ways.

First, Jiutian must provide more value than a conventional inspection stack. Natural-language instructions and adaptable planning sound attractive, but a buyer will ask which additional tasks become economical or safer.

Second, Ruoyu must preserve the reliability of specialized automation while adding model-driven decisions. If a conventional robot follows a fixed route successfully, greater autonomy only matters when it handles meaningful variation without creating unacceptable risk.

Huawei represents another part of this competitive context. Its energy strategy emphasizes cloud, edge, connectivity, data platforms, and industry models. Huawei’s smart oilfield material describes a cloud-edge-device architecture supporting production sites and station operations.

Huawei says its smart field system can connect operational data and support AI applications across energy facilities. Those figures remain company claims, but the architecture illustrates what a robot supplier must integrate with.

Jiutian cannot operate as an isolated intelligence appliance. An industrial robot needs authenticated access to facility data, defined permissions, reliable communications, event logging, and a path for human intervention.

China Mobile and other telecommunications providers occupy another layer. Private wireless networks and edge infrastructure can determine whether robots receive timely data or continue operating safely during a connection failure.

This creates a crowded technical stack. A robot manufacturer controls motors, joints, sensors, and protective hardware. Ruoyu supplies perception and planning intelligence. A cloud or platform vendor manages data services. A network operator provides communications. The energy company remains responsible for the facility.

Each interface can become a failure point. It can also become a commercial boundary where vendors compete for control.

Ruoyu’s opportunity lies between rigid automation and broad enterprise infrastructure. If Jiutian can translate operating goals into dependable robot actions, it could become a reusable layer across different machines.

That opportunity also explains why the company emphasizes cross-body deployment. Energy operators use wheeled platforms, quadrupeds, manipulators, drones, and specialized service robots. Supporting several forms could reduce the cost of rebuilding intelligence for every machine.

However, a common brain does not eliminate hardware differences. A quadruped crossing a trench has different constraints from a manipulator opening a fuel cap. Sensors, payloads, reachable spaces, braking behavior, and protective certifications remain body-specific.

The winning architecture will probably combine shared models with tightly constrained local controllers. The model can interpret context and propose actions. Certified control layers must enforce speed, force, position, and emergency limits.

Ruoyu now has to show that Jiutian fits this layered structure. Energy operators already have robots and digital platforms. They do not need another AI demonstration unless it expands useful autonomy without weakening operational control.

Jiutian’s Mechanism Connects Perception, Planning, and Motion

Jiutian’s main technical argument is that one model-driven system can adapt a robot’s plan as the physical environment changes.

Ruoyu says Jiutian combines embodied perception, embodied planning, and embodied execution. Embodied AI refers to intelligence that receives information from a physical environment and acts through a machine inside that environment.

The company’s Jiutian specifications describe support for video, audio, text, point clouds, and force data. Point clouds are three-dimensional collections of measured positions, often generated by depth cameras or lidar.

This input mix can help a robot answer several distinct questions. It must identify relevant objects, estimate their positions, understand the operator’s goal, and determine which surfaces permit interaction.

The company says Jiutian can decompose an abstract assignment into smaller actions. It can then refine the sequence and generate a trajectory for the robot’s end effector, such as a gripper or tool.

That mechanism targets a genuine limitation of traditional industrial robotics. Many machines perform reliably only after engineers program routes, fixtures, coordinates, and exception rules for a stable environment.

Energy operations are rarely perfectly stable. Equipment positions vary. Lighting changes. Outdoor surfaces collect dust or water. Workers and vehicles enter the scene. A valve may resist movement differently after wear or temperature changes.

A model that interprets those variations could reduce programming effort. It could also let one robot handle several related jobs instead of repeating one fixed motion.

Ruoyu’s filling-station example shows why this is difficult. The company says its Lanyue 01 robot can open a vehicle’s fuel door, remove a cap, retrieve a nozzle, dispense fuel, return the nozzle, and close the vehicle.

That sequence contains more than object recognition. The robot must distinguish vehicle designs, locate movable parts, estimate contact forces, verify each completed step, and stop safely when conditions depart from the expected state.

A fuel cap that appears open may remain partially engaged. A nozzle may enter at the wrong angle. A driver may move near the robot. A vehicle may not match the training distribution.

Long sequences amplify small errors. If each step depends on the previous step’s physical result, an unnoticed failure can corrupt the rest of the plan.

Force sensing is particularly relevant. Vision may show that a gripper reached a handle, while force data reveals that the grasp failed. The robot needs both signals before proceeding.

Ruoyu says Jiutian uses closed-loop correction. A closed loop compares the intended outcome with fresh sensor information, then adjusts the next action. This differs from executing an entire sequence without checking results.

The company also describes world-model functions. A world model estimates how the environment will change after an action. For example, it might predict the expected position of a cap after the robot rotates it.

These capabilities remain company-reported. Ruoyu has not published a peer-reviewed evaluation covering task completion, intervention rates, abnormal conditions, and recovery across multiple energy sites.

The available technical specifications also reveal a useful architectural choice. Jiutian can run on an edge device with local computing, rather than relying entirely on a remote cloud.

Local processing can reduce latency and preserve some functions during network problems. It also supports tighter control over sensitive facility data.

However, edge deployment introduces its own constraints. Models must fit within available memory and computing capacity. Thermal limits, power consumption, software updates, and cybersecurity become part of the robot’s operating profile.

Industrial buyers will also want to know which decisions occur inside the model and which remain under deterministic control. A language-based planner should not directly override an emergency stop, safety interlock, or hazardous-zone restriction.

This is where Jiutian’s “brain” metaphor becomes incomplete. A deployable industrial system needs more than cognition. It needs certified hardware, bounded controllers, monitoring, audit trails, maintenance procedures, and trained operators.

NIST makes a similar distinction in its work on physical AI. The agency notes a persistent gap between embodied AI research and systems that manufacturers can implement in real operations.

Ruoyu’s mechanism addresses the intelligence side of that gap. Its commercial future depends on whether the company can also support the surrounding engineering discipline.

A successful pilot should therefore measure more than whether Jiutian finishes a staged task. It should record completion rates, human interventions, false alarms, unsafe proposals, latency, recovery time, and performance after environmental changes.

The crucial question is not whether Jiutian can plan. It is whether the complete robot system knows when the plan should stop.

Hazardous Work Turns Every AI Claim Into a Safety Claim

In an energy facility, adaptability has value only when operators can bound, observe, and reverse the robot’s behavior.

Ruoyu has reported that Lanyue 01 entered public trial operation at a filling station in April 2026. The company also says the robot received an explosion-protection certification for its intended environment.

Those statements deserve attention, but they do not independently verify the complete autonomous system. Hardware certification and AI behavior assurance solve different problems.

Explosion protection addresses whether equipment can operate without creating an ignition source under specified hazardous conditions. It does not prove that a perception model recognizes every object or that a planner selects the correct action.

Software reliability also cannot compensate for unsuitable hardware. A correct plan still becomes dangerous if a sensor fails, a seal degrades, or an actuator produces unexpected force.

The opposite is equally true. Certified mechanical and electrical components do not guarantee that an adaptive model will behave correctly under unfamiliar conditions.

This separation should shape how energy companies evaluate Jiutian. They need a layered safety case, which is a structured body of evidence connecting hazards, controls, tests, and operating limits.

At the model layer, evaluators should test perception under glare, darkness, dust, occlusion, vibration, and sensor degradation. Energy sites contain reflective metal, steam, narrow structures, and repetitive equipment that can confuse visual systems.

At the planning layer, testers should introduce incomplete instructions, contradictory data, and changing conditions. The system should reject unsafe requests and request human guidance when confidence falls below an accepted threshold.

At the control layer, the robot must enforce hard limits that the planner cannot bypass. Those limits can include restricted zones, maximum force, permitted tool orientation, and emergency stopping conditions.

At the operational layer, facilities need logs that reconstruct what the robot sensed, decided, and executed. Without that record, teams cannot investigate near misses or improve procedures.

Cybersecurity cuts through every layer. A connected robot receives software updates, reads facility information, and may transmit video or equipment data. Compromised credentials or manipulated sensor inputs can turn a useful machine into an operational risk.

Data governance also matters because training and field logs may expose facility layouts, asset conditions, employee movements, or proprietary procedures. Buyers will want clear rules governing storage, access, retention, and model improvement.

A recent Scientific Reports study on robotic hazard data describes one reason safety validation remains difficult. Researchers cannot freely reproduce dangerous industrial events in operating facilities.

Synthetic data and simulation can expand coverage, but neither perfectly captures real equipment, weather, sensor noise, or human behavior. Field evidence remains essential.

Ruoyu therefore faces a verification problem shared by the entire embodied AI sector. The situations that matter most are rare, hazardous, and difficult to stage repeatedly.

The company can reduce that uncertainty through bounded deployment. An early robot might inspect assets and flag anomalies while leaving physical intervention to people. Later versions could perform manipulation under supervision.

This staged approach would not diminish Jiutian’s ambition. It would create the operational data needed to support greater autonomy.

Energy operators should also separate uptime from autonomy. A robot might complete many patrols but still require frequent remote assistance. Another might operate independently yet cover only a narrow, carefully prepared route.

Useful disclosure should state both figures. Buyers need to know how often the robot completes a mission and how often a person intervenes.

Similar clarity is needed around exceptions. Average task performance can conceal rare failures with severe consequences. Evaluation should emphasize worst-case behavior and safe recovery, not only mean completion time.

Human factors deserve equal attention. Workers must understand when the robot is operating, what area it controls, how to stop it, and who holds authority during an abnormal event.

An adaptive robot can also change work routines. Operators may start relying on automated readings, while maintenance staff inherit new responsibilities for sensors, models, networks, and software versions.

This organizational burden can determine whether a pilot becomes a production system. A capable robot that requires constant specialist support may not improve total operating economics.

Ruoyu’s claims should be judged with this broader frame. Jiutian does not need to achieve unrestricted autonomy to become useful. It needs to automate a valuable set of tasks within limits that operators can understand and enforce.

Until independently reported trials provide those details, its performance in hazardous energy settings remains promising but unproven.

Three Signals Will Show Whether Jiutian Can Leave the Conference Stage

The next meaningful evidence will come from operating records, repeatable deployment, and integration with established energy systems.

The first signal is a named, sustained field deployment with measurable results. Ruoyu has described filling-station trials, but buyers need reporting that identifies the task scope, operating duration, intervention rate, and safety record.

A deployment becomes more convincing when it continues through ordinary weather, equipment variation, maintenance cycles, and customer behavior. A staged demonstration cannot expose the same distribution of conditions.

Evidence from several locations would carry more weight than success at one prepared site. Repeated installations show whether the system transfers without months of custom engineering.

This signal would strengthen Ruoyu’s case if operators disclose consistent completion and low intervention rates. It would weaken the case if every site requires extensive retraining, physical modification, or continuous remote control.

The second signal is third-party validation of the complete system. That assessment should cover model behavior, control boundaries, cybersecurity, hardware protection, and human operating procedures.

A component certificate is useful but insufficient. The robot brain, sensors, actuators, network connections, and facility interfaces form one operating system in practice.

Independent testing should also publish the conditions under which the system must stop or hand control to a person. Safe refusal is a capability, especially when a robot encounters an unfamiliar vehicle or ambiguous equipment state.

This signal would strengthen the company’s argument if test results define clear performance boundaries and show reliable fallback behavior. It would weaken it if verification remains limited to promotional demonstrations or isolated components.

The third signal is integration with an energy operator’s existing data and control environment. Jiutian needs to exchange information with inspection platforms, maintenance records, work-order systems, and site controls without becoming an uncontrolled automation island.

That integration should enforce identity, authorization, and logging. A robot should receive only the information and permissions required for its assigned task.

It should also preserve local safety during a network outage. Losing a cloud service or wireless connection must not leave the machine in an undefined physical state.

This signal would strengthen Ruoyu’s position if an operator uses Jiutian across different robot bodies through a common, governed interface. It would weaken the thesis if each deployment remains a stand-alone showcase.

The Beijing conference placed Ruoyu in the right room, but access to major energy companies is only the beginning. These organizations already understand the value of remote inspection and hazardous-duty automation.

Their harder question is whether model-driven autonomy improves operations enough to justify new verification, cybersecurity, integration, and maintenance work.

Jiutian’s strongest idea is not a humanoid shape or a conversational interface. It is a reusable intelligence layer that might let specialized robots adapt to more tasks.

Its greatest risk follows from the same idea. A reusable brain can spread one model’s weaknesses across multiple bodies and facilities unless every deployment has effective local safeguards.

Ruoyu can resolve that tension with evidence. It needs sustained operating data, independent system testing, and integrations that survive ordinary industrial constraints.

For developers and enterprise buyers, the practical lesson is straightforward. Follow deployments rather than conference appearances. Ask how often humans intervene, which decisions remain deterministic, and what happens after a sensor or network failure.

The next Jiutian announcement will matter most if it answers those questions. Until then, Ruoyu has earned a place in China’s energy robotics conversation, but not yet a verdict on production readiness.