# What is physical AI in oil and gas?

> Physical AI in oil and gas is AI that senses a site and acts in it through a machine, such as a drone, crawler, underwater vehicle or edge camera, not only on documents and data. Because its mistakes move hardware in hazardous places, it is governed like new plant equipment, with explosion protection, OT security and human oversight.

- Label: Fact
- Type: Explainer
- Published: 2026-11-15T00:16:49+09:00
- By: Jun Zhong
- Researched and drafted by: Beacon's AI newsroom
- Checked by: Jun Zhong on 2026-11-15
- Topics: Methane and emissions, Regulation
- Canonical: https://beacon.winda.ai/library/physical-ai/

On the seabed, a remotely operated vehicle (ROV) finds a pipeline by sonar, switches to a camera to line up with a bolted flange, grips it and taps it to hear whether the bolts are loose. A University of Houston project funded by the US Bureau of Safety and Environmental Enforcement (BSEE) ran that sequence end to end without a pilot steering each move, and its models judged bolt tightness with more than 90% accuracy in most setups, on flanges from 4.25 to 21 inches [1]. That is physical AI: the model's output is a movement of hardware, not a paragraph of text. This page explains what the term covers, how it differs from AI on documents, and which rules apply when it works on an oil and gas site.

## What does "physical AI" mean?

The term is young and used loosely. At an April 2026 hearing, the chairman of a US House science subcommittee described "physical AI" as the integration of AI into robotic systems [2]. The US National Institute of Standards and Technology (NIST) runs a project named Physical AI and Data Generation for Robotics, which builds metrics, test methods and datasets for AI-enhanced robots and says a large gap remains between embodied AI in research and what manufacturers can actually deploy [3].

Researchers more often say embodied AI. A 2024 survey describes it as the bridge between cyberspace and the physical world, with four research targets: perception, interaction, the agent itself and moving skills learned in simulation to real machines [4]. A 2025 survey, by contrast, uses "physical AI" for models that reason about physical laws and predict how the world will change [5]. No standards body has fixed a definition, so this page uses the robotics sense.

## What are the parts of a physical AI system?

The BSEE subsea project shows each part in one machine [1].

| Part | What it does | In the subsea flange project |
|---|---|---|
| Sensing | Collects raw signals from the site | Scanning sonar for long range, cameras up close, an inertial unit for position [1] |
| Perception | Turns signals into objects the machine can act on | Pipeline lines pulled out of sonar returns; a vision model that locates the flange [1] |
| Planning and control | Decides the next move and holds position | Marker-based station keeping in drifting water, then an approach and grasp [1] |
| Actuation | Acts on the equipment | A gripper fitted with paired piezoceramic transducers that send stress waves through the joint [1] |
| Judgement | Classifies what it measured | Machine learning models that grade each bolt as tight or loose [1] |

On fixed plant, the same chain can sit in a box beside the process. Joint guidance from the US Cybersecurity and Infrastructure Security Agency (CISA) and partner agencies, December 2025, notes that some modern programmable logic controllers (PLCs) and edge controllers run small pre-trained models for local anomaly detection and for holding a known safe state [6].

## How is it different from AI on documents and data?

| Question | AI on documents and data | Physical AI |
|---|---|---|
| Where it usually runs | Large language models at the business levels of the Purdue model (4 and 5), often on data exported from the plant [6] | Predictive models in the operational levels (0 to 3), next to sensors and actuators [6] |
| What it acts on | Records, drafts and recommendations [6] | Machines and the process; operational technology (OT) controls physical systems that can harm people or property [6] |
| How it is tested | Against known answers before use [7] | Also in the field, because results from a lab or controlled setting can differ from risks in real operation [7] |
| What catches a failure | A reviewer and the change process [6] | Failsafes that fall back to conventional automation or manual operation, built into functional safety and incident response [6] |

The two often meet in one job: an [AI agent](/library/industrial-ai-agent/) drafts an inspection on the records side, and a crawler or drone carries it out. CISA's guidance spans both ends, from language models on exported data to models running in field devices and controllers [6]. Errors on the records side are covered in [why general AI tools get engineering answers wrong](/library/ai-errors-in-engineering/).

## Where is physical AI used at oil and gas sites?

- **Tall structures and tanks.** Flare stacks, chimneys, cooling towers and tanks have been inspected from scaffolds, ropes or helicopters; a 2018 review from Khalifa University in Abu Dhabi notes that drone inspection still needed a certified pilot and an inspection engineer, and names automated flight as the research goal [8]. Image models can then mark corrosion in drone photographs [9]. See [drone inspection](/library/drone-inspection/) and [AI in inspection](/library/ai-inspection/).
- **Pipelines that pigs cannot pass.** A PHMSA-supported program developed two robots for live gas transmission lines: one carrying a remote field eddy current sensor for 6 and 8 inch pipe, one carrying a magnetic flux leakage sensor for 20 to 26 inch pipe [10]. See [robotic crawlers](/library/robotic-crawlers/).
- **Subsea joints.** The BSEE project above aimed to cut the need for divers at bolted flanges [1].
- **Methane leaks.** Under the 2024 US methane rule for new sources, an operator may replace some optical gas imaging surveys with periodic screening or continuous monitoring by an approved methane detection technology, at a frequency set by its detection threshold [11]. A confirmed detection still has to be followed up on the ground and repaired [11]. See [leak detection and repair](/library/leak-detection-and-repair/).
- **Inside tanks and vessels.** US rules list tanks and vessels as typical confined spaces and count an entry as soon as any part of a worker's body passes the plane of the opening [12].

**Analysis:**
The confined space definition is why remote tools matter in O&G inspection: it is written around a person's body, so a camera or crawler that goes in while the crew stays outside changes what the permit has to control [12]. It does not remove the hazards at the opening, and the machine itself becomes equipment that must suit the atmosphere inside.

## What rules apply to a robot or drone on site?

| Constraint | What it asks | Source |
|---|---|---|
| Hazardous (classified) locations | Electric equipment must be intrinsically safe, approved for the location, or shown safe for it; approved equipment is marked with class, group and operating temperature; zone classification is supervised by a qualified registered professional engineer | [13] |
| Equipment for explosive atmospheres | IEC 60079-0:2026, edition 8.0 of June 2026, sets the general requirements for building, testing and marking equipment used in or with explosive atmospheres | [14] |
| Airspace, for drones in the US | The pilot or an observer keeps the drone in sight; flights stay under 400 feet above ground unless within 400 feet of a structure; the FAA can waive line of sight | [15] |
| The OT network | Connect AI like any new OT system, through approved paths and the existing demilitarized zone; keep an inventory of AI components; log their inputs and outputs | [6] |
| Change on offshore facilities | Written management of change for new equipment, procedures and operating conditions, covering technical basis, safety and environmental impact, and approval before use | [16] |
| AI risk | Safety risks that could cause serious injury or death get the most urgent priority and the most thorough risk management | [7] |

A robot's motors, batteries and computers are electric equipment, so the first two rows apply to the machine as a whole, not only to its sensors [13]. Securing the network side is covered in [OT cybersecurity](/library/ot-cybersecurity/).

**Analysis:**
What makes physical AI new is where an error lands. A wrong answer on a document waits for a reviewer; a wrong move by a machine happens before anyone reads a log, which is why the CISA guidance warns that problems with active AI can grow before operators notice [6]. The practical test for any proposal is whether the machine, its network link and its fallback have each passed the same approvals a new piece of plant equipment would.

## A common mix-up

Physical AI is often confused with automation and with digital twins. CISA puts ladder logic routines and a class of safety instrumented systems among programmed algorithms, and sets them apart from machine learning, which learns from data and can drift when the process changes; where AI influences control, it asks for failsafes and points where a person can step in [6]. A [digital twin](/library/digital-twin/) is a virtual model fed by the asset; physical AI is the machine on the asset, and it may feed a twin or take its instructions from one.

## Sources

1. [TAP 812, Robotic Assistive Smart Touch Inspection of Offshore Oil and Gas Pipelines (project page and final research abstract), US Bureau of Safety and Environmental Enforcement (BSEE), Technology Assessment Program; research by the University of Houston](https://www.bsee.gov/robotic-assistive-smart-touch-inspection-of-offshore-oil-and-gas-pipelines), 2026-02-28
2. [Opening Statement of Chairman Jay Obernolte at Robots Made in America: Advancing U.S. Leadership in Manufacturing and Automation, US House of Representatives, Committee on Science, Space, and Technology](https://science.house.gov/2026/4/opening-statement-of-chairman-jay-obernolte-at-robots-made-in-america-advancing-u-s-leadership-in-manufacturing-and-automation), 2026-04-21
3. [Physical AI and Data Generation for Robotics (project page), US National Institute of Standards and Technology (NIST), Engineering Laboratory](https://www.nist.gov/programs-projects/physical-ai-and-data-generation-robotics)
4. [Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI (Y. Liu, W. Chen, Y. Bai and others), arXiv:2407.06886, arXiv (preprint)](https://arxiv.org/abs/2407.06886), 2024-07-09
5. [Aligning Perception, Reasoning, Modeling and Interaction: A Survey on Physical AI (K. Xiang, T. J. Zhang, Y. Huang and others), arXiv:2510.04978, arXiv (preprint)](https://arxiv.org/abs/2510.04978), 2025-10-06
6. [Principles for the Secure Integration of Artificial Intelligence in Operational Technology (joint guidance, TLP:CLEAR), US Cybersecurity and Infrastructure Security Agency (CISA), Australian Signals Directorate's Australian Cyber Security Centre and partner agencies](https://www.cisa.gov/sites/default/files/2026-01/joint-guidance-principles-for-the-secure-integration-of-artificial-intelligence-in-operational-technology-508cV2.pdf), 2025-12-03
7. [Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, US National Institute of Standards and Technology (NIST)](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf), 2023-01-26
8. [Inspection of vertical structures in oil and gas industry: A review of current scenario and future trends (V. Sudevan, A. Shukla, H. Karki), SEG/AAPG/EAGE/SPE Research and Development Petroleum Conference, Abu Dhabi, pp. 65-68, DOI 10.1190/RDP2018-41952188.1, Khalifa University of Science and Technology (research portal abstract)](https://khazna.ku.ac.ae/en/publications/inspection-of-vertical-structures-in-oil-and-gas-industry-a-revie/), 2018-05-09
9. [CorrDetector: A Framework for Structural Corrosion Detection from Drone Images using Ensemble Deep Learning (A. R. M. Forkan et al.), arXiv:2102.04686; published in Expert Systems with Applications, arXiv (preprint record with journal DOI 10.1016/j.eswa.2021.116461)](https://arxiv.org/abs/2102.04686)
10. [Completion of Development of Robotics Systems for Inspecting Unpiggable Transmission Pipelines (project 353, public abstract), US Pipeline and Hazardous Materials Safety Administration (PHMSA), Pipeline Safety Research and Development Program](https://primis.phmsa.dot.gov/rd/projects/353)
11. [40 CFR 60.5398b, alternative GHG and VOC standards for fugitive emissions components affected facilities when using an alternative technology, eCFR, US Office of the Federal Register (US Environmental Protection Agency rule)](https://www.ecfr.gov/current/title-40/chapter-I/subchapter-C/part-60/subpart-OOOOb/section-60.5398b)
12. [29 CFR 1910.146, Permit-required confined spaces, paragraph (b) definitions, eCFR, US Office of the Federal Register](https://www.ecfr.gov/current/title-29/section-1910.146)
13. [29 CFR 1910.307, Hazardous (classified) locations, eCFR, US Office of the Federal Register](https://www.ecfr.gov/current/title-29/section-1910.307)
14. [IEC 60079-0:2026, Explosive atmospheres, Part 0, Equipment, General requirements, edition 8.0 (publisher's scope), International Electrotechnical Commission (IEC), webstore](https://webstore.iec.ch/en/publication/114218), 2026-06-16
15. [14 CFR Part 107, Small Unmanned Aircraft Systems (sections 107.31, 107.51 and 107.205), eCFR, US Office of the Federal Register](https://www.ecfr.gov/current/title-14/chapter-I/subchapter-F/part-107)
16. [30 CFR 250.1912, What criteria for management of change must my SEMS program meet?, eCFR, US Office of the Federal Register](https://www.ecfr.gov/current/title-30/chapter-II/subchapter-B/part-250/subpart-S/section-250.1912)

## From Winda

Winda supplies, installs and services the probes, holders, fittings and injection quills that feed the agents. Every reading arrives with its source, so the analysis starts from measured metal, and each reading ends in a decision an engineer signs. [See Winda's field services](https://winda.ai/services/field-services-instruments/)

Published by Winda Beacon (https://beacon.winda.ai).
