Physical AI

When AI Steps Out of the Digital World, and What It Needs on the Ground

AI Digest

by Csaba Fekszi

2026. 10. 06.
8 minutes

AI is moving into robots, vehicles, and factory equipment that sense, decide, and act. Some of it pays off at scale today, much of it is still a pilot, and the most useful first step is often a camera.

Has physical AI had its ChatGPT moment yet?

NVIDIA thinks so. At CES 2025, the company’s press release quoted its chief executive, Jensen Huang: “The ChatGPT moment for robotics is coming.” A year later, its CES 2026 release quoted him again: “The ChatGPT moment for robotics is here.”

Gartner reads the same market more soberly. Its Hype Cycle for Physical AI, published in July 2026, calls the field transformational and concludes that its hype and risks currently run ahead of business results.

Both views can hold at once. This article explains what physical AI is, where it already pays off, and where it is still a pilot. It also shows what a manufacturing or logistics company can prepare now, well before the first robot arrives.

What is physical AI

Gartner lists physical AI among its top strategic technology trends for 2026. It describes the field as AI inside machines and devices that sense, decide, and act in the real world, from robots and drones to smart factory equipment.

Factories have used robots for decades. The International Federation of Robotics counted a record 5 million industrial robots in operation worldwide in 2025, after more than 600,000 new installations during the year.

Most of them repeat movements that engineers programmed in advance. Physical AI adds perception and judgment, so a robot can locate a part lying at an odd angle, notice a person in its path, and adjust its plan. The federation itself names advances in artificial intelligence, machine vision, and sensing as the drivers expanding what robots can do.

Three abilities define the field, and each one runs on data. Sensing turns the physical world into images and signals. Deciding interprets them, often with models trained on video and simulation. Acting moves an arm, a vehicle, or a conveyor, and every action changes the scene the system has to read next.

Where physical AI already pays off

The clearest results so far come from large fleets of task-specific machines in controlled environments. Amazon reported in mid-2025 that it had deployed its one millionth robot. It also introduced DeepFleet, a generative AI foundation model designed to coordinate how those robots move. By Amazon’s own figures, it cuts their travel time by 10%.

Two details behind the headline matter for any company. Amazon built DeepFleet on its own extensive data about how inventory moves inside its sites, and its robotics program goes back to 2012. At its next-generation fulfillment center in Shreveport, Louisiana, the advanced robotics require 30% more employees in reliability, maintenance, and engineering roles.

The ingredients are instructive: a narrow task, a structured environment, deep operational data, and new skills to keep the machines running. Gartner makes the same point about skills, noting that physical AI needs people who can bridge IT, operations, and engineering.

Where physical AI is still a pilot

Humanoid robots draw the most attention, and they sit furthest from production. Gartner predicts that through 2028, under 100 companies will take humanoid proofs of concept beyond experimentation. Under 20 will run humanoids in live production for manufacturing and supply chain use.

Abdil Tunca, a supply chain analyst at Gartner, sums up the gap:

“The promise of humanoid robots is compelling, but the reality is that the technology remains immature and far from meeting expectations for versatility and cost-effectiveness.”

Abdil Tunca, Senior Principal Analyst, Gartner, January 2026

Its analysts name four barriers: limited dexterity and adaptability in unstructured environments, difficult integration with existing systems and workflows, high costs, and short battery life. It sees a better fit for most companies in polyfunctional robots: multi-purpose machines whose shape follows the task. One example is a wheeled robot with a telescopic arm that moves boxes, picks cases, and scans inventory.

With today’s technology and prices, Gartner estimates that a humanoid costs several times as much as a task-specific polyfunctional robot, while delivering lower throughput and uptime. Its advice is outcome-driven automation aimed at specific bottlenecks.

A pilot in Kecskemét

Hungary has hosted one of these pilots. Mercedes-Benz tested Apptronik’s Apollo humanoids at its Kecskemét plant and at its Digital Factory Campus in Berlin, on tasks such as moving components to the production line and carrying out quality checks.

When the carmaker took a stake in Apptronik in March 2025, Reuters reported that a handful of the robots had been trained for specific tasks through teleoperation. People operated them remotely, and the robots learned each task to repeat it on their own.

That is what a sensible pilot looks like: a few machines, clear tasks, and people teaching the robot.

Pick the bottleneck first, then the machine.

What physical AI needs from your company

A robot on the shop floor needs four things beyond its own sensors. It needs the order it is working on, the part and its tolerances, the right storage location, and the rule for when something goes wrong.

Those answers live in ERP, MES, and warehouse management systems, and in the heads of experienced operators. Even NVIDIA’s CES 2025 release noted that physical AI models are expensive to build and depend on large volumes of real-world data and testing.

Gartner names limited real-world data as a key constraint on physical AI. It advises AI leaders to streamline their sensor and OT/IoT platforms, the operational technology that runs machines and collects their signals.

This is where an older discipline comes back in. Master data, clean interfaces, and clear ownership of each data source decide whether a robot reads the same reality as the ERP. At Omnit, we have spent more than 15 years integrating enterprise systems. Physical AI raises the same questions with higher stakes, because a wrong record can now move a physical object.

Physical AI depends on the same data foundations as other enterprise AI systems: clear ownership, consistent definitions, and trusted data across systems. We examine that foundation in Why Do AI Projects Fail Without Data Governance →.

Start by seeing

For many companies, the first physical AI project only watches. Cameras are often already installed over production lines and loading bays. AI image and video analysis can turn their footage into events: a missing part, a skipped step, a person in a restricted zone.

Our own early activity recognition pilot, run on recorded footage from an operational environment, showed both the potential and the limits. Distinct, consistent movements were easier to recognize, while subtle or overlapping ones proved harder. Camera position and lighting noticeably changed the results. Even partial recognition revealed recurring work patterns that stay invisible in conventional operational data.

The lesson carries over to robots. A machine that will act on what it sees should first prove, on your own footage, that it sees correctly.

Teach the system to see before you let it act.

New rules for AI in machinery from January 2027

Regulation is moving at the same time. From 20 January 2027, the EU Machinery Regulation replaces the 2006 Machinery Directive. Safety components and embedded safety systems whose behavior evolves through machine learning fall into its highest-risk category. That category requires an independent notified body to assess conformity before the machine reaches the market.

The rules also reach companies that change their own lines. Under Article 18, whoever substantially modifies a machine can be treated as its manufacturer for that modification. Before you retrofit an AI safety function onto an existing line, ask your safety and legal teams whether the change counts as substantial.

The AI Act adds a second layer later. Following an amendment the Council approved in June 2026, its high-risk rules for AI embedded in products apply from 2 August 2028. For machinery, the Commission can add requirements through the Machinery Regulation itself.

The same preparation applies to AI projects more broadly: define the process, data, ownership, and success criteria before choosing the technology. We cover that framework in The Key Steps to a Successful AI Implementation →.

Five steps before your first physical AI project

1. Name one bottleneck.
Pick the process step that limits throughput, quality, or safety, and agree on a metric to show improvement, such as scrap rate or waiting time per shift.

2. See before you act.
Put cameras or sensors on that step and measure what really happens in your own footage, against go-or-no-go criteria agreed in advance.

3. Connect events to records.
Link every event to an order, a part, and a location from ERP, MES, or the warehouse system, so each alert lands in a process with a named owner.

4. Choose the machine last.
Compare a task-specific or polyfunctional robot with a humanoid on cost per completed task, uptime, and the maintenance skills each one needs.

5. Check the safety route early.
Decide who signs off safety functions, whether a planned change counts as substantial, and which conformity assessment applies from January 2027.

The ChatGPT moment that counts is your own

Has physical AI had its ChatGPT moment, then? In NVIDIA’s press releases, it has. Gartner’s analysts still see hype and risk running ahead of the business results.

For your company, the real moment arrives when the groundwork is in place: a mapped process, clean master data, and cameras that already turn footage into evidence. Companies that build this now can choose their robots on their own terms, whenever the market’s moment comes.

Start with one bottleneck, one camera, and clean data. The robot can follow.

Is physical AI worth starting now, and where?

Physical AI is moving fast. For your company, the useful question is whether to act now, and on which process. In a free 30-minute consultation, a senior Omnit expert helps you find the logical next step, and an honest not yet is one possible answer. If the first step is to see more of your operation, our image and video analysis service starts with a focused pilot on your existing cameras.

Sources

Amazon. (2025). Amazon launches a new AI foundation model to power its robotic fleet and deploys its 1 millionth robot. Read article →
Council of the European Union. (2026). Artificial Intelligence: Council gives final green light to simplify and streamline rules. Read article →
Gartner. (2025). Gartner Identifies the Top Strategic Technology Trends for 2026. Read article →
Gartner. (2026). Gartner Predicts Fewer Than 20 Companies Will Scale Humanoid Robots for Manufacturing and Supply Chain to Production Stage by 2028. Read article →
Gartner. (2026). Leverage Data and World Models to Advance Physical AI. Read article →
Gartner. (2026). Hype Cycle for Physical AI, 2026. Read article →
International Federation of Robotics. (2026). Five Million Robots Now Operate in Factories Globally. Read article →
Murphy, B. (2026). Machine Head. Read article →
NVIDIA. (2025). NVIDIA Launches Cosmos World Foundation Model Platform to Accelerate Physical AI Development. Read article →
NVIDIA. (2026). NVIDIA Releases New Physical AI Models as Global Partners Unveil Next-Generation Robots. Read article →
Omnit. (2026). When Activity Recognition Meets Reality. Read article →
Waldersee, V. (2025). Mercedes-Benz Tests Apptronik Robots in Factories. Read article →

Csaba Fekszi

Csaba Fekszi is an IT expert with more than two decades of experience in data engineering, system architecture, and AI-driven process optimization. His work focuses on designing scalable solutions that deliver measurable business value.