Physical AI combines edge computing, digital twins, and OT/IT integration to bring artificial intelligence directly into production processes. In this article, we share the Revelis vision for making Physical AI operational through PlugAIn, alongside a concrete case study, Aquila project within the AI-REGIO framework, where Physical AI applied to quality control reduced inspection times by 60%.
Physical AI is artificial intelligence acting right where production happens: on the assembly line, inside the electrical panel, and on the machinery, not just in back offices. For a manufacturing company, this means deploying AI close to factory data (at the edge), enabling it to interface with a digital twin of the product or process, and overcoming friction between OT and IT systems. Let’s dive into the details.
What Is Physical AI (and why is everyone talking about It now?)
Physical AI represents the new frontier of artificial intelligence: no longer just systems that respond to text prompts in a chat, but systems that perceive, interpret, and act within the physical world through cameras, sensors, PLCs, and industrial machinery. Recent years have seen significant buzz around generative AI and agentic AI applied to text, documents, and back-office workflows. Physical AI applies this exact principle, giving AI the capacity to act, rather than merely answer, to the factory floor, where decisions confront physical realities: mechanical tolerances, vibrations, lighting conditions, and component wear.
For a manufacturer, Physical AI is far from an abstract concept. It is the distinction between a model that identifies a defect in a manually uploaded photo and a system that detects that same defect in real time on the line as the part passes under a camera. This shift, from manually loaded data to field-captured data, is the core of Physical AI.
From the cloud to the factory floor: why AI needs to be near the machine
Until now, most enterprise AI applications have resided in the cloud: language models, dashboards, and analytical tools accessible via a web browser. Physical AI changes the rules of the game because it must respond within the timeframes of the physical process it observes, milliseconds for a robotic arm, seconds for a visual inspection. It cannot afford the latency of round-trip communications with a remote data center.
Added to this are practical constraints unique to industrial environments: factory floor connectivity is rarely as reliable as office Wi-Fi, production data is often sensitive or bound by customer non-disclosure agreements, and streaming continuous high-definition video feeds from dozens of cameras to the cloud is cost-prohibitive. Consequently, industrial Physical AI relies heavily on edge computing: processing occurs as close to the data source as possible, while the cloud remains reserved for heavy model training, historical data aggregation, and multi-facility analytics.
The digital twin: The “common language” between design and reality
A critical second component of Physical AI is the digital twin: a dynamic digital representation of a product, facility, or process that is continuously updated to mirror physical reality. It is not a static simulation, but a live baseline against which Physical AI measures deviations: comparing what should be happening according to design specifications with what is actually taking place inside the machine, panel, or plant.
Revelis has long developed solutions in this domain. With Ei-TWIN, for example, we applied the digital twin paradigm to the energy sector, integrating real-time IoT data to optimize consumption and energy exchange within a pharmaceutical plant and an energy community. Physical AI generalizes this principle to broader industrial contexts: wherever a design exists (a CAD file, a schematic, a specification sheet) alongside a physical build, automated comparison between the two yields actionable value.
OT/IT integration: where friction arises (and gets solved)
Implementing Physical AI on the factory floor means tackling a long-standing industrial challenge: the gap between the OT world (Operational Technology, PLCs, SCADA systems, sensors, line cameras) and the IT world (ERP systems, data warehouses, cloud applications). These two domains operate with different protocols, lifecycles, and cultures, historically communicating poorly.
To function, Physical AI must act as a bridge between these layers: capturing and interpreting real-time OT data, normalizing it, and serving it in a format that IT systems (and personnel) can use for decision-making. We addressed this challenge through a different angle in the InCoP project, where environmental and biometric field data collected at a bakery (for Target S.p.A.) were processed using machine learning and reasoning to optimize shift scheduling. There too, value stemmed from enabling field devices to communicate with management systems. In Physical AI applied to quality and production, that same principle connects cameras, PLCs, and CAD files.
MHS: an emerging standard for connecting AI to physical hardware
The challenge of machine protocol heterogeneity is significant enough that on August 27, 2026, Anthropic released a research preview of the Model Hardware Standard (MHS): a specification designed to connect AI agents to physical laboratory and production instruments, aiming to reduce integration times from weeks or months down to hours or minutes.
Technically, MHS introduces a standardized driver layer between the operating system and the hardware device, built on simple read/write primitives (e.g., “read temperature”, “set temperature”), automatic device discovery, and natural language tags documenting capabilities and safety bounds. It is designed to be model-agnostic and accessible via standard protocols like the Model Context Protocol (MCP), command line interfaces, or APIs.
Initial results shared by Anthropic alongside research partners demonstrate its potential:
- QuEra Computing (quantum computing): An AI agent reduced laser recalibration times from 150 seconds to 6 seconds while boosting the success rate from 58% to 99.3%.
- Carnegie Mellon University: Integration of dilution series for dose-response curve tracking went from weeks to roughly 8 hours, executing three times faster.
- Genentech: An agent automated a BCA assay for protease analysis, autonomously optimizing liquid flow parameters and identifying errors caused by air bubbles.
- University of Washington: The system remotely coordinated qPCR and instrument monitoring from a unified dashboard, featuring real-time tracking of amplification curves.
- HHMI Janelia Research Campus: Unified seven proprietary software programs for two-photon zebrafish microscopy, enabling real-time data analysis during acquisition.
- Tetsuwan Scientific: Managed a qPCR workflow for water pollution monitoring at San Pedro Creek, achieving 12% higher predictive accuracy than the manufacturer’s published specifications.
These milestones confirm that Physical AI, when backed by a unified hardware standard, can drastically reduce integration times across diverse domains, from pharmaceutical research to environmental monitoring.
However, existing limitations remain clearly defined: Anthropic notes that Claude lacks reliable “physical intuition”—struggling to predict collisions, inertia, or structural instability. High-risk decisions still require human approval, and safety roadmaps (hard stops, emergency shutoffs, error handling) are actively being refined prior to any open-source release. This aligns directly with the core philosophy behind Physical AI at Revelis: AI accelerates execution, but human oversight remains an essential layer. MHS serves as a clear indicator of industry direction, representing a potential abstraction layer for the OT gateways Revelis builds when integrating new machinery into PlugAIn.
PlugAIn: how Revelis operationalizes physical AI on the shop floor
To make Physical AI practical for manufacturing companies, Revelis relies on PlugAIn, our proprietary platform that models production processes using BPM (Business Process Management) and Big Data Analytics components. The core approach involves collecting and processing data generated by machinery, sensors, and line cameras to deliver actionable intelligence to decision-makers—whether an algorithm or a human operator—integrating edge processing and a comparison engine between design (the digital twin) and field reality.
PlugAIn combines three complementary layers of Physical AI:
- Computer Vision to perceive the physical state of components.
- Reasoning Engine to compare that state against the expected design model.
- Orchestration Layer to decide, based on the output, whether to proceed automatically or involve a human operator.
This structural principle, AI proposes, humans decide on high-value tasks, mirrors Revelis’s agentic AI solutions on the IT side, here extended to the physical environment of the factory floor.
A real-world case study: Aquila, physical AI for quality control (AI-REGIO Project)
The best way to understand the impact of Physical AI is through a real-world application. Within the European project AI-REGIO (Horizon 2020, grant agreement No. 952003), Revelis developed Aquila, an Automatic Quality Inspection system validated at Elettrocablaggi, an Italian manufacturer of electrical panels. It serves as a direct example of Physical AI: cameras observing physical products, a CAD model representing design specifications, and an automated engine comparing the two worlds.
Before Aquila, quality control for electrical panels was entirely manual. An operator visually cross-referenced each manufactured panel against engineering schematics, taking an average of 340 seconds per panel with an error rate of roughly 1 in every 250 inspected panels, typical metrics for a repetitive, labor-intensive process prone to human fatigue. This is precisely where thoughtfully applied Physical AI creates impact.
The Aquila solution combines two complementary AI techniques:
- Computer Vision: A Mask R-CNN object detection neural network identifies installed components and their precise spatial positioning from photos of physical panels.
- Symbolic Reasoning: The CAD design file is transformed into a logical representation, effectively the electrical panel’s digital twin. An automated reasoning engine based on Answer Set Programming (the DLV engine) compares the two graph representations (the vision-captured state vs. the target design) and generates a compliance report flagging missing, extraneous, or misaligned components.
Results recorded during testing highlight the power of quality-focused Physical AI:
- The object detection model achieved an Area Under the Precision-Recall Curve (AUC) of 0.947, indicating high reliability.
- Average inspection time dropped by 60%.
- Error detection accuracy improved by 30% compared to a standalone computer vision approach.
- The system advanced from Technology Readiness Level TRL 4 to TRL 6, marking a major step toward full industrial deployment.
Why physical AI matters for manufacturing SMEs
The Aquila project demonstrates that Physical AI is not an academic exercise, but a tangible competitive lever for small and medium-sized manufacturing enterprises (SMEs):
- Less time spent on repetitive inspection yields higher throughput using existing headcounts.
- Fewer errors translate directly to fewer returns, claims, and rework cycles.
- A digitized process generates an audit trail of data, imagery, and results, transforming daily operations into an intelligence asset over time.
Once validated on quality control for electrical panels, this framework extends naturally to any operational context involving design-versus-reality verification: predictive maintenance, assembly verification, or surface and weld inspection. Once introduced to a single line item, Physical AI scales value across the enterprise.
This goes beyond theory: Revelis actively applies Physical AI to live machinery monitoring. At Target, an IoT sensor network connected to PlugAIn monitors every production phase in an industrial bakery, from ingredient dosing to baking, identifying machine idle times to inform preventive maintenance schedules.
In another predictive maintenance project, sensors mounted on textile loom feeders stream data via MQTT to a recurrent neural network (GRU). The network predicts warp yarn breaks with 70% accuracy (AUC 0.67), providing advance warnings that allow operators to intervene before machine stoppage occurs. These implementations confirm that Physical AI delivers value equally well in direct machinery monitoring and quality control.
The Revelis vision in brief
Physical AI is not simply another cloud platform promise; it is the strategic choice to bridge artificial intelligence with the physical reality of production using edge computing, digital twins, and robust OT/IT integration. Through PlugAIn, Revelis consolidates insights gained across projects like Aquila (AI-REGIO), Target, Ei-TWIN, and InCoP to make Physical AI a practical, measurable pathway for small and mid-sized manufacturers. It is our hardware-adjacent expression of the core principle guiding Revelis: making AI actionable.
Mini-Glossary for Non-Technical Readers
| Physical AI | Artificial intelligence systems that perceive and act within the physical world (via sensors, cameras, and machinery) rather than operating exclusively on text or digital data. |
| Edge AI / Edge Computing | Processing data close to its physical source (on the machine or local factory network) rather than in a remote data center, reducing response latency and network dependencies. |
| Digital Twin | A dynamic digital representation of a physical product, facility, or process synchronized with its real-world state. |
| OT (Operational Technology) | Hardware and software that monitor and control physical factory equipment (PLCs, SCADA systems, line sensors). |
| IT (Information Technology) | Enterprise data processing systems (ERPs, data warehouses, cloud apps) that must interface with OT layers. |
| Computer Vision | AI techniques focused on extracting meaningful information from digital images and video feeds. |
| Answer Set Programming (ASP) | A symbolic AI framework based on formal logic rules, used to verify whether a given state complies with defined rules. |
| TRL (Technology Readiness Level) | A scale from 1 to 9 assessing technology maturity, from basic research (TRL 1) to full commercial deployment (TRL 9). |
| MHS (Model Hardware Standard) | A specification introduced by Anthropic (August 2026 research preview) to connect AI agents to laboratory and manufacturing hardware using standardized drivers. |
| MCP (Model Context Protocol) | An open protocol developed by Anthropic that enables AI agents to connect to external tools and data sources; MHS extends this concept to physical device control. |
Ready to Evaluate Physical AI for Your Facility?
Revelis guides manufacturing enterprises through every step of implementing Physical AI on the factory floor, from edge computing and digital twins to full OT/IT integration.
Learn more at www.revelis.eu · Contact us at info@revelis.eu
Authors: Massimiliano Ruffolo and Achille Abritta
