Physical AI in Energy and Utilities
Article authored by my friend Imran Dar, Industry Executive Director at Oracle
"Consider a simple calculation. If a task requires 50 independent steps, each with a 99% success probability, the chance of completing every step without failure is about 60.5%. If every failed step can be detected and safely retried once, and each retry is independent with the same success probability, completion rises to about 99.5%. "
Why physical AI matters for energy
Energy companies have spent decades automating physical work. They have invested in process control, drilling automation, remote operations and equipment that can respond to changing conditions. A useful discussion of physical AI has to recognise that starting point. Otherwise, we risk presenting achievements the industry already knows as something entirely new.
I see the opportunity in extending what these systems can do, particularly where work is variable, access is difficult or people still have to interpret a situation before a machine can act. Better perception and learning could make more of that work practical to automate. The value will depend on the operating problem, the reliability of the equipment and the cost of deploying it.
There is a second opportunity in connecting physical work to the rest of the enterprise. A robot may identify a deteriorating bearing earlier than a scheduled inspection. Turning that finding into a better outcome still requires someone to assess the consequence, secure a part, organise the work and confirm the repair. These decisions involve maintenance, operations, supply chain and finance. When information or coordination is holding up the response, better integration can make a material difference.
Enterprise AI could improve that response further by helping people prioritise work, consider alternatives and handle exceptions. That is a proposition to test. Connecting a robot to a maintenance system establishes integration; showing that AI improves the resulting decisions requires additional evidence. I would keep those two investment decisions separate.
This paper examines what energy companies have already achieved, what is changing in the underlying technology and how leaders can assess the next investment. Its examples cover oil and gas, oilfield services, electric networks, solar construction, gas distribution and adjacent process industries. Water, nuclear and conventional generation require further sector-specific evidence. The aim is to make the opportunity understandable without losing sight of how energy operations actually work.
The industry is already well into automation
The distinction between established automation and physical AI is less tidy than much of the current discussion suggests. Industrial control has long included feedback, optimisation and adaptation. Some of today's applications already use learned models. A new label should help explain an additional capability rather than erase that history.
For this paper, physical AI means learned intelligence that interprets physical conditions and helps determine or execute a physical action. The system might advise an operator, inspect equipment or control a specific process. Its actual authority matters more than the label. Broader industry definitions sometimes include simulation and other supporting technologies, which is one reason physical AI market estimates are difficult to compare. [S28, S32]
General-purpose physical AI is a more demanding ambition. It aims to reuse a model across more tasks, environments or types of machine, with less engineering for each new application. Vision-language-action research is pursuing that ambition. It does not follow that an energy company needs a general-purpose model for every useful application. A specialist inspection model or an established controller may do the job well. [S11–S13]
| Capability |
Role in the operation |
Example |
| Established automation |
Runs engineered sequences, feedback control and optimisation. |
Equipment sequencing and grid restoration. |
| Specialist industrial AI |
Learns a defined perception, prediction or control task. |
Defect detection or learned process control. |
| Embodied autonomy |
Combines sensing, navigation or manipulation with task execution. |
Inspection robots and installation equipment. |
| Foundation model physical AI |
Reuses learned representations across a wider range of instructions or tasks. |
A VLA adapted to new manipulation work. |
These capabilities overlap. A single installation may combine all four, with people supervising some activities and directly controlling others. The investment question is whether learning handles enough additional variability, or reduces enough engineering effort, to justify its cost.
The industry's operating record provides a useful benchmark. Nabors reported an unmanned rig floor on an ExxonMobil test well in 2021. It also said that crew size remained similar, with people moving into supervision, maintenance and other work. The achievement was a change in how the work was performed and where people were exposed. The companies did not publish the test-pad performance results. [S03]
SLB subsequently reported 99% autonomous control of a 2.6-kilometre well section on Equinor's Peregrino C platform. Its separate claim of a 60% improvement in rate of penetration covered a five-well programme. Those are different measures, and neither should be read as a 60% reduction in total well cost. Chevron and Halliburton's 2025 announcement describes another form of adaptation: intelligent fracturing in Colorado that adjusts completion behaviour using subsurface feedback and domain algorithms. The companies described expected efficiency benefits without publishing a measured result. [S04–S05]
Electric networks offer an equally important comparison. Duke Energy Florida estimated that its self-healing technology avoided more than 280,000 extended customer outages in 2025. It reported coverage of about 82% of its roughly two million customers. Automatic fault detection and rerouting already deliver a substantial operating capability. That disclosure does not establish the use of a foundation model. Any new proposal should be compared with the reliability and economics of the automation already in place. [S06]
What is changing now
Several developments make this worth executive attention. Models are becoming better at relating visual observations to instructions and actions. Simulation offers more ways to train and test behaviour. Suppliers are investing in the computing and software needed to put these capabilities into machines. Together, these developments could reduce the effort required to automate work that has resisted conventional approaches. [S11–S15]
The commercial interest is visible. Reuters reported Arm's creation of a Physical AI unit in January 2026 and an announced Skild AI deployment on Foxconn assembly lines in March. The existing equipment base is also substantial: the International Federation of Robotics reported 542,000 industrial robot installations in 2024. That figure includes conventional industrial robots, but it indicates the engineering and integration base available to support further development. [S36–S38]
Governments are paying attention for their own industrial and strategic reasons. The US Department of Energy's 2026 Genesis Mission challenges include autonomous laboratories and advanced manufacturing. The European Commission's Apply AI Strategy includes energy and robotics. China's 2025 Government Work Report summary identifies embodied AI among future industries. These initiatives point to sustained interest in capability development, although commercial performance still has to be established application by application. [S33–S35]
For energy leaders, this arrives alongside growing demand for the infrastructure that supports AI. The IEA's April 2026 central projection puts global data-centre electricity consumption at 950 TWh in 2030, compared with 485 TWh in 2025. These figures cover all data centres. The same agency's work on AI for energy describes opportunities in maintenance, grid operations and industrial efficiency. Energy companies therefore face both an infrastructure demand challenge and an opportunity to improve how their own assets operate. [S01–S02]
The major advisers help frame the management questions. McKinsey discusses adaptation, economics and changes to human work; BCG distinguishes the maturity of different robotic capabilities. Accenture connects physical and agentic AI with changes in operating processes. EY emphasises the continuing investment needed to scale, PwC examines responsibility and governance, and Deloitte takes a broad view of the convergence of AI and robotics. I use these perspectives to inform the questions an executive should ask. Specific performance claims in this paper rest on the deployment or research source concerned. [S27–S32]
The technology needs a closer look
The deep-tech work matters because physical systems have to deal with timing, forces, uncertainty and wear. A model that produces a sensible explanation has demonstrated only part of what an industrial machine needs. The machine must also identify the correct object, act within its physical limits and respond appropriately when conditions change.
Understanding the physical situation
An inspection system may combine camera images, thermal readings, acoustic signals and position data. These measurements have different strengths. A visible mark might indicate surface damage; a thermal pattern might suggest overheating; sound might help identify a leak. The useful result is an interpretation tied to the right asset and its operating condition. ANYmal's inspection applications illustrate this use of multiple sensing methods. [S09, S23]
Site conditions can change that interpretation. Steam may obscure a camera. Dust, poor lighting or a moved piece of equipment may confuse localisation. Sensor drift can make a familiar reading unreliable. The response should be designed with the task: take another measurement, ask for review, stop or return to a known operating mode. A confidence score only helps if testing establishes how well it reflects uncertainty in those conditions.
Learning actions and transferring skills
Vision-language-action models, usually shortened to VLAs, connect visual observations and instructions to robot actions. RT-2 explored how knowledge learned from web data could support robotic control. OpenVLA released a generalist model and methods for adapting it. The π0.5 research demonstrated broader generalisation using different kinds of training data. For an operator, the attraction is the prospect of teaching a new task with less engineering and fewer demonstrations. The task still needs to be validated where it will be performed. [S11–S13]
Learning can take several forms. Imitation learning uses demonstrations of desired behaviour. Reinforcement learning improves a policy against an objective through interaction, often in simulation. Defining that objective is an engineering responsibility: speed is of limited value if the policy damages equipment or compromises product quality. Learned perception can also work alongside a physics-based model and established control logic. There is no requirement for a single model to make every decision. [S10, S12–S13]
Planning and physical execution may run in different systems. Google DeepMind's Gemini Robotics 1.5 announcement describes an embodied reasoning model working alongside an action model. That separation is relevant to enterprise integration: a system can help decide what should be done while another executes a specific task. Each still needs limits appropriate to its role. [S14]
Simulation and evidence from the real asset
A world model predicts aspects of how an environment may evolve. A physics simulator calculates behaviour under specified assumptions. A digital twin represents a particular asset or process, potentially updated with operating data. They can work together, but a convincing generated scene is not evidence that forces, fluid behaviour or heat transfer are represented accurately.
NVIDIA's Cosmos 3 announcement describes combining world generation, physical reasoning and action prediction. The potential benefit is a wider range of training and testing situations. For industrial use, the relevant test is whether those situations preserve the physical properties that determine success or failure. Calibration against measurements remains necessary. [S15]
National-laboratory grid research offers a useful precedent. NREL reported hierarchical-control experiments involving about one million simulated nodes connected to real power hardware. This was a large simulation and hardware-in-the-loop exercise, rather than a million-device live deployment. It shows how controlled testing can examine scale before an operator takes the corresponding risk in service. [S18]
Computing and mechanical reliability
The deployed model has to fit the machine's response time, power supply, thermal limits and connectivity. Sending every decision to a remote service may be impractical. Compressing a model for local execution can change performance. OpenVLA's 2024 experiments showed that inference speed affected control results even when offline token accuracy looked similar. The relevant unit of testing is the deployed machine and its software together. [S12]
Mechanical design can be just as decisive. Reach, traction, payload, battery endurance and corrosion resistance determine which jobs a machine can perform and how often people must attend to it. The machine also has to suit the site's environmental and hazardous-area requirements. A humanoid is one possible form. A drone, fixed sensor, wheeled robot or existing automated machine may be a more practical choice for the work.
Reliability becomes harder to establish as tasks get longer. Published research identifies problems such as near-miss grasps, early releases and drift across sequences of actions. Safety research also examines attacks on multimodal inputs and the difficulty of intervening in time. These problems require testing of failure detection and recovery, as well as successful task completion. [S16–S17]
Consider a simple calculation. If a task requires 50 independent steps, each with a 99% success probability, the chance of completing every task without failure is about 60.5%. If every failed step can be detected and safely retried once, and each retry is independent with the same success probability, completion rises to about 99.5%. These are illustrative assumptions, not observed robot performance. Real failures can be related, and a retry may be unsafe or ineffective. The example shows why recovery deserves as much attention as an isolated demonstration of a skill.
For deep-tech investors, I would look closely at access to representative operating data, test facilities, dependable hardware and evidence of repeatable results. A foundation model can accelerate development. The work required to make it dependable at a customer site may determine the commercial advantage.
What the deployment evidence shows
The clearest cases involve specific jobs. They demonstrate useful progress across inspection, construction and control, with different levels of autonomy. They also make it possible to discuss physical AI in terms of operating work rather than a distant vision of a fully autonomous enterprise.
National Grid announced routine deployment of centrally operated autonomous drones for transmission inspection in England and Wales in September 2025. The inspection data informs maintenance and investment programmes, with people continuing to operate and support the service. In solar construction, AES describes Maximo's use of AI vision and automated module installation, alongside human assistance for activities including replenishment and movement between rows. A March 2026 company release reported 100 MW installed at its Bellefield project using four robots, beyond the nearly 10 MW described on an earlier AES page. [S07–S08, S44]
Process control provides an example of learning that predates the current foundation-model discussion. Yokogawa and JSR reported a 35-day reinforcement-learning control test on a distillation application in 2022. In March 2023, Yokogawa announced formal adoption at the same plant, then operated by ENEOS Materials, following almost a year of control experience. This moves the evidence beyond a short demonstration while keeping the scope clear: it concerns a particular distillation application. [S10, S40]
Remote facilities present a different operating need. ANYbotics describes ANYmal deployment at the Equinor-operated Northern Lights carbon-storage facility, connected to Equinor's Flotilla platform. Equinor's broader robotics work includes subsea inspection and intervention. A resident drone at Njord completed a reported 240-day dive ending in January 2025. Remaining subsea for that period is a meaningful engineering achievement; it is a measure of residence rather than 240 days of wholly autonomous task execution. [S09, S39, S43]
Gas distribution adds useful experience in working inside live infrastructure. ULC Technologies reported that CISBOT had sealed 50,000 joints in live cast-iron gas mains by March 2023, including work for Con Edison and National Grid. It is teleoperated, so its relevance here is the physical task and deployment experience. In China, a 2024 conference paper reports that State Grid had deployed more than 3,000 substation inspection robots by 2022. These cases broaden the evidence beyond the more frequently discussed offshore and drilling examples. [S45–S46]
I would take two lessons from this record. First, there are worthwhile applications well before a machine can operate across an unrestricted range of tasks. Second, the measures need to match the claim. An inspection completed, a joint sealed, a well section drilled and a customer outage avoided are different outcomes. Grouping them under physical AI does not make their economics or reliability interchangeable.
Connecting the physical work to the enterprise
The enterprise connection becomes easier to understand through an ordinary maintenance problem. Suppose a robot detects an abnormal thermal pattern on a compressor bearing. This is an illustrative example. Before the finding is useful, it needs the correct equipment identity, operating context and enough evidence for a reliability assessment. The response then depends on the consequence of failure and the available opportunity to intervene.
The maintenance planner needs to know whether the part is available and whether qualified people can do the work. Operations has to agree the isolation and timing. The repair must be recorded, and subsequent readings should establish whether it resolved the condition. Finance needs a defensible account of the cost and any avoided loss. An accurate observation is the beginning of that work.
I use five functions to organise this connection. They are a discussion framework, informed by established enterprise and control integration principles, rather than a proposed industry standard. ISA-95 already addresses enterprise and control integration. The table makes the responsibilities relevant to physical AI explicit without attempting to replace that standard or prescribe a security architecture. [S19]
| Function |
Responsibility |
| Enterprise records and results |
Maintain asset, work, inventory and financial records; establish what was completed and what it achieved. |
| Business decisions and coordination |
Assess consequences, plan the response and route approvals or exceptions. |
| Physical intelligence |
Interpret measurements and plan or perform a permitted task. |
| Local control and edge systems |
Execute control, enforce operating limits and provide fallback behaviour. |
| Assets and environment |
Provide measurements and the physical setting in which work is performed. |
A mobile inspection robot may run its own navigation and control without commanding the plant. An AI application in drilling or process control has a more direct relationship with operating equipment. Both need clear permissions, but their engineering and assurance requirements will differ. Safety, cybersecurity and change control apply across the full arrangement.
The information moving into enterprise systems needs a reliable asset identifier, location, timestamp and measurement history. Instructions moving back need an approved task and clear operating limits. Integration must also cope with ordinary failures: a message arrives twice, connectivity is lost, an inventory record is stale or a work order is rejected. The system should distinguish a recommendation from an approved order and an attempted repair from a completed one.
Enterprise AI could help when these decisions involve changing context. It might assemble evidence for a planner, compare intervention options or flag a conflict between an asset risk and an outage schedule. Transaction rules should still enforce authority, budgets and valid asset relationships. Where a stable rule already handles the decision well, conventional integration may be sufficient.
There are public examples of the connection taking shape. SAP describes dispatching work orders to ANYmal through Field Service Management and returning inspection results to business workflows. Its March 2026 article also describes AI-agent involvement. ANYbotics' October 2026 Shift announcement brings together fleet operations, inspection analytics and maintenance integration. These sources show how physical work can connect to enterprise records. They do not provide a controlled comparison of the additional value from enterprise AI. [S23–S24]
Oracle's release 25D documentation illustrates a separate distinction. The Maintenance Work Order Builder is a conversational agent for users that creates, retrieves and updates work orders; it excludes material and resource transactions. The sibling Postmaintenance Work Recorder supports those transactions. Machine integration uses services such as the maintenance work-order REST APIs, including an action for creating condition-based work orders. A robot workflow would need to be designed and validated around the relevant interfaces. These are descriptions of the documented 25D capabilities, not an assessment of the latest release or an end-to-end robotics offering. [S25, S41–S42]
Accenture's Physical AI Orchestrator and its wider work on systemic AI explore related connections in manufacturing. The underlying idea therefore has precedents. My interest is how to apply it to energy operations, where asset risk, work authority and the response process determine whether the connection creates value. [S26, S29]
Most companies will make this work across a mixture of industrial systems, asset management, ERP, data platforms and cloud services. The useful starting point is the installed environment. Reliable interfaces and clear ownership of records matter more than selecting one supplier for every function.
Making the investment case
The business case should identify the constraint before pricing the solution. An inspection robot may reduce difficult site visits. Automated installation may increase construction throughput. Better coordination may reduce the delay between a confirmed fault and its repair. Each proposition has a different baseline and a different way of earning a return.
| Application |
Measures that connect to the operating result |
| Inspection and maintenance |
Verified defects closed, missed defects, time to repair, repeat visits and total cost. |
| Drilling and completions |
Total well economics with comparable geology and scope; human exposure and execution consistency. |
| Grid inspection and repair |
Useful defect coverage, repair time and service reliability. |
| Solar construction |
Accepted modules per crew-shift, rework and actual effect on the project schedule. |
| Process optimisation |
Quality-adjusted output and energy use under comparable conditions. |
The cost must cover the working service: equipment, site preparation, integration, communications, computing, supervision, maintenance and recovery when things go wrong. Data preparation, simulation, training and model validation also continue after installation. EY's emphasis on the ongoing investment needed to scale is relevant here. Buying the machine is only part of the commitment. [S30]
An invented inspection example shows why I would assess the investments separately. Assume a physical AI capability costs $900,000 upfront and $250,000 a year to operate. It removes $150,000 a year of external inspection and travel spending. Each avoided production-loss hour is worth $25,000 in contribution margin, and the capability avoids eight hours a year with the existing response process. Annual net benefit is then $100,000, giving a nine-year simple payback.
Now assume conventional integration and process changes cost a further $300,000 upfront and $60,000 a year, and avoid another ten lost hours. Finally, assume enterprise AI costs an additional $200,000 upfront and $120,000 a year, and avoids two more hours. The table uses seven years of level annual cash flows discounted at 10%, with upfront spending at the start.
| Investment step |
Added upfront cost |
Added annual cost |
Added avoided hours per year |
Annual net benefit |
Seven year NPV |
| Physical AI with the existing process |
$900,000 |
$250,000 |
8 |
$100,000 |
−$0.41m |
| Conventional integration and process changes |
$300,000 |
$60,000 |
10 |
$190,000 |
+$0.63m |
| Enterprise AI added to that process |
$200,000 |
$120,000 |
2 |
−$70,000 |
−$0.54m |
Under these assumptions, the physical capability and conventional integration together create about $0.21 million of net present value. Adding enterprise AI reduces the combined value to about minus $0.33 million. The AI step needs approximately 6.44 additional avoided hours a year to break even, compared with the two assumed. This is a test of the method using invented inputs. It is neither a forecast nor a judgement about the economics of an actual product. Taxes, financing, replacement expenditure and residual value are excluded.
The example also shows why a successful combined project can conceal an uneconomic addition. A team should be able to explain which step creates each benefit and which spending is required to obtain it. Avoided downtime and additional production value should not count the same gain twice. Redeployed employee time becomes a cash saving only when spending falls; otherwise, the case needs to show the valuable work that the released capacity enables.
For utilities, the result may appear through service quality, available capacity, justified capital deferral or operating efficiency. The regulatory and commercial arrangements determine how customers and shareholders benefit. For oilfield services, faster execution can have different consequences under a time-based contract and a performance-based contract. Value creation and value capture need to be considered together.
I would test the enterprise AI proposition with four conditions where the operating setting permits: the existing workflow; physical AI with the existing response process; physical AI with conventional integration and a redesigned response process; and that same redesigned process with enterprise AI added. Comparable assets or a staggered rollout can help, provided the evaluation accounts for other changes, operating exposure and differences between sites.
The comparison between the last two conditions estimates the incremental contribution of enterprise AI when other material conditions are comparable. Integration and process redesign should already be present in both. This makes the evidence more useful than comparing a new, fully redesigned service with an old process and crediting every improvement to AI.
Operating responsibility and the longer term
As physical and enterprise systems become more connected, a mistaken observation can travel further. It may create unnecessary work, reserve the wrong part or influence an operating decision. DOE's assessment of AI in critical energy infrastructure highlights risks from incomplete data and operating outside training conditions. NIST's OT security guidance explains why reliability, timing and safety shape the design of industrial systems. These considerations belong in the deployment design. [S20, S22]
The permitted operating conditions and actions should be explicit, with a response for conditions the system cannot handle. Existing protective functions must retain their intended role. In process industries, the ISA/IEC 61511 safety lifecycle is a relevant established reference; the responsible engineering and safety teams need to assess the actual application and changes. Adding a model does not itself establish compliance. [S21]
Human responsibility needs equal clarity. Someone must have the authority and practical ability to approve a mission, stop an operation, authorise a repair and release a changed system into service. A person nominally supervising several machines may have too little time or information to intervene effectively. The arrangements need to work during an exception, when attention is already under pressure. [S31]
Learning from operating experience should feed a controlled release process. Preserve the data, model version and event history needed to understand a decision. Test updates before deployment and retain a way to restore a known configuration. Procurement should address support, data rights, integration responsibilities and what happens when a supplier changes a model or withdraws a service.
The operating owner should remain accountable for the outcome, supported by field personnel, controls and OT specialists, AI engineers, enterprise-process owners and security teams. Nabors' account of changing crew duties is a useful reminder that automation can redistribute work without immediately reducing headcount. People will still need to handle exceptions, maintain equipment and retain the practical competence needed when automation is unavailable. [S03]
Looking ahead, I expect progress to be uneven. Through 2028, the most credible expansion is in specific inspection, assistance and adaptive-control tasks. Between 2029 and 2032, a more consequential test will be whether companies can reuse skills and integrations across fleets and sites at lower cost. Broader coordination across asset operations and business functions could develop further through the 2030s. These are working scenarios from an October 2026 starting point; the dates are less useful than the evidence that would support them.
That evidence should include fewer interventions under representative conditions, lower deployment effort at the second and third sites, and a sustained improvement in completed work. If integration remains expensive or operating conditions vary too much, specialist automation with human decision support may retain the stronger case. Strategy should accommodate that outcome.
The longer-term implications go beyond the robot. Operating data and validated procedures may become more reusable across an enterprise. Service contracts may place greater weight on verified results, requiring agreement on how outcomes and responsibility are measured. Suppliers that connect physical operations to business decisions may gain influence over both. Operators will need to protect access to their data and their ability to change providers.
There is also a practical limit to better detection. More findings can increase the maintenance backlog if the organisation cannot respond. Better planning may use crews and outage windows more effectively, but it cannot make a missing part immediately available or create specialist capacity overnight. The physical AI programme must be considered alongside the resources needed to act on its output.
Where I would start
I would begin with one recurring operating problem, an accountable owner and a credible baseline. Inventory the automation already in place. Identify where the work loses time, incurs avoidable cost or exposes people to risk. Then determine whether the missing capability is better sensing, physical execution, integration or decision support. That assessment may lead to a robot, a learned controller or a simpler process improvement.
The first phase should establish the operating and financial hypothesis, the permitted authority and the conditions that would cause the team to stop or change direction. The next should test the complete response using simulation, replay or observation-only operation where appropriate. Asset identity, approvals, failed transactions and the workload reaching the maintenance team deserve attention alongside model performance.
Only then should the team move to supervised operation within agreed limits. The time required depends on the asset and its assurance requirements; a fixed 90-day calendar is a poor substitute for sufficient operating exposure. The decision to expand should rest on verified results and a credible account of what the next site can reuse.
My view is that physical AI deserves serious attention from energy leaders because it could extend what machines can reliably do in the field. Connecting that capability to enterprise decisions could make the improvement more valuable where coordination is a real constraint. The case for enterprise AI becomes stronger when it demonstrably helps the organisation complete better work at an acceptable cost. That is the evidence I would want before committing to scale.
Research scope and notes
This is an independent working paper based on public research, operator and supplier disclosures, agency publications, journalism and advisory reports. Deployment claims remain attributed to the reporting organisation. Research demonstrations, product descriptions and policy announcements serve different purposes and have not been treated as equivalent evidence of commercial performance. The framework, investment method and long-term scenarios are my analysis.
The examples are selective rather than a global sector census. They include activity in the United States, the United Kingdom, Norway, Brazil, Japan and China. They do not establish equivalent readiness across water, nuclear, conventional generation or markets not examined here. Private executive discussions have not been quoted or presented as independently verified evidence.
The source register preserves the 46 identifiers used in v02. Source access and review were uneven: some material was available only as publisher excerpts, abstracts or company statements reproduced by another publication. The independent reviews did not verify every source in full. Product descriptions retain their cited release or announcement dates. The illustrative economics use invented assumptions, and the paper has not been peer reviewed.
References
Source identifiers are retained from v02 for comparison with the research register. Links lead to the cited documents. S43 and S45 are trade press accounts; S44 is a reprint of a company release. Listed dates describe the source, not a claim that all material was fully verified in this editorial revision.
[S01] International Energy Agency. 16 April 2026.
Key Questions on Energy and AI Executive summary
[S02] International Energy Agency. 10 April 2025.
Energy and AI AI for energy optimisation and innovation
[S03] Nabors. 7 October 2021.
World’s First Fully Automated Land Rig Has Successfully Drilled Its First Well
[S04] SLB. 30 January 2024.
SLB and Equinor Drill Most Autonomous Well Section To Date
[S05] Halliburton and Chevron. 12 June 2025.
Chevron and Halliburton enable intelligent hydraulic fracturing
[S06] Duke Energy Florida. 6 February 2026.
Smart self healing technology investments help keep customers lights on
[S07] National Grid. 4 September 2025.
National Grid Rolls Out Centralised Autonomous Drone Inspections across its network
[S08] AES. Undated page accessed 6 October 2026.
Maximo AI Solar Robot for Clean Energy
[S09] ANYbotics. 23 June 2025.
ANYmal deployed at Northern Lights CCS Facility
[S10] Yokogawa and JSR. 22 March 2022.
AI autonomously controls a chemical plant for 35 consecutive days
[S11] Brohan and colleagues. 28 July 2023 preprint.
RT-2 Vision-Language-Action Models Transfer Web Knowledge to Robotic Control
[S12] Kim and colleagues. 2024 paper version dated 5 September.
OpenVLA: An Open-Source Vision-Language-Action Model
[S13] Black and colleagues. CoRL 2025 PMLR volume 305 pages 17 to 40.
π0.5: A Vision-Language-Action Model with Open World Generalization
[S14] Google DeepMind. September 2025.
Gemini Robotics 1.5 brings AI agents into the physical world
[S15] NVIDIA. 31 May 2026.
NVIDIA Launches Cosmos 3 the Open Frontier Foundation Model for Physical AI
[S16] Zhang Qi and Zheng. 14 November 2025 preprint.
Experiences from Benchmarking Vision Language Action Models for Robotic Manipulation
[S17] Li and colleagues. 26 April 2026 preprint.
Vision Language Action Safety Threats Challenges Evaluations and Mechanisms
[S18] National Renewable Energy Laboratory now hosted by NLR. 2021.
Large Scale Experiments Demonstrate Advanced Controls for Autonomous Energy Systems
[S19] International Society of Automation. Current overview accessed 6 October 2026.
ISA-95 Series of Standards Enterprise Control System Integration
[S20] NIST Stouffer and colleagues. 28 September 2023.
Guide to Operational Technology Security SP 800-82 Revision 3
[S21] International Society of Automation. Current overview accessed 6 October 2026.
ISA-84 Series of Standards
[S22] US Department of Energy CESER. 26 April 2024.
Potential Benefits and Risks of Artificial Intelligence for Critical Energy Infrastructure
[S23] SAP Alexa MacDonald. 30 March 2026.
How ANYbotics and SAP turn industrial inspections into business insights
[S24] ANYbotics. 1 October 2026.
Launch of Shift for robot fleet operations and industrial inspection
[S25] Oracle. Release 25D documentation accessed 6 October 2026.
AI Agent Maintenance Work Order Builder 25D
[S26] Accenture. 28 October 2025.
Physical AI Orchestrator for software defined facilities
[S27] McKinsey. 24 June 2026.
The age of thinking machines Perspectives on the future of robotics
[S28] Boston Consulting Group. 14 April 2026.
How Physical AI Is Reshaping Robotics Today and What Comes Next
[S29] Accenture. 21 April 2026.
Systemic AI at the root of manufacturing performance
[S30] EY. March 2026.
Five strategic questions for the C suite on physical AI
[S31] PwC. 17 July 2026.
Redefining governance in the age of physical AI
[S32] Deloitte Insights. Tech Trends 2026 edition.
AI goes physical Navigating the convergence of AI and robotics Tech Trends 2026
[S33] US Department of Energy. 12 February 2026.
26 Genesis Mission Science and Technology Challenges
[S34] European Commission. 2025 strategy current policy page accessed 6 October 2026.
Apply AI Strategy
[S35] State Council of China official portal. 6 March 2025.
Chinese robots show skills
[S36] Reuters. 7 January 2026.
Arm launches Physical AI unit joining rush to robotics
[S37] Reuters Akash Sriram. 16 March 2026.
Skild AI Nvidia deploy robot brain on Blackwell assembly lines
[S38] International Federation of Robotics. 25 September 2025.
Global Robot Demand in Factories Doubles Over 10 Years
[S39] Equinor. Undated page accessed 6 October 2026.
Drones and robots in Equinor
[S40] Yokogawa. 30 March 2023.
Yokogawa’s Autonomous Control AI Is Officially Adopted for Use at an ENEOS Materials Chemical Plant
[S41] Oracle. Release 25D documentation accessed 7 October 2026.
AI Agent: Postmaintenance Work Recorder
[S42] Oracle. Release 25D REST API documentation accessed 7 October 2026.
Maintenance Work Orders: create condition-based work orders
[S43] Offshore magazine. 28 January 2025.
Saipem and Equinor complete record-breaking autonomous mission
[S44] Maximo and AES, wire release reprinted by Barchart. 25 March 2026.
Maximo completes 100 MW of robotic solar installation
[S45] Robotics 24/7. March 2023.
ULC Technologies’ CISBOT Seals 50,000 Joints in Live Gas Mains
[S46] State Grid Intelligence Technology authors, ACM conference paper.
State of the Art and Development Trends for Inspection Robots Applied in Substations
| 
