What is physical AI? What it means on a UK warehouse floor

What physical AI means: software with a body

Physical AI, sometimes written as embodied AI, covers AI systems that perceive, reason about and act within the physical world, particularly at human scale (4). These systems combine AI models with sensors, control systems, actuators and machines such as robots or autonomous vehicles (4). The emphasis is on the complete process: perceiving an environment, planning an action, then physically executing the task to do work (4). That is what separates it from digital AI or generative AI, which mostly stays inside information (4).

The boundaries of the term are not standardised. It is often treated as a continuation of earlier research in robotics and embodied intelligence rather than a separate new field (4). In practice you will see it applied to self-driving vehicles, industrial and warehouse robots, humanoid and service robots, drones and delivery robots. It also covers autonomous agricultural, construction and mining machines (4). Household versions already exist: robotic vacuum cleaners and lawn mowers that use sensors and navigation software with limited human control (4). If you want the vocabulary unpacked, our plain-English AI glossary covers the surrounding terms.

The phrase got its current prominence during the AI boom of the 2020s, as development expanded from digital applications toward humanoid robots, self-driving vehicles and smart factories (4). Money followed the vocabulary. Nvidia’s chief executive described the physical AI business as already driving nearly $10 billion in annual revenue for the company, during an appearance on the All-In Podcast in March 2026 (3). He began highlighting physical AI as Nvidia’s second-most important growth category in 2025 (3).

One point worth settling early: physical AI is not a chatbot attached to a machine. It is the loop of sensing, deciding and doing that runs on the machine itself. It runs close enough to real time that a missed decision has consequences you can see on the floor.

Why it matters to a UK warehouse or factory

The short answer for an owner or operations lead is that this technology has stopped being a lab curiosity and is now being bought. An industrial software executive in a recent MIT Technology Review discussion said one study suggests almost a 78 per cent rise in AI adoption across the industrial sector over two years (5). The study is not named, so treat the figure as a signal rather than a measurement (5). Hardware money is moving too: Nvidia has bet billions of dollars on physical AI technologies such as robotics and self-driving cars (3).

For a warehouse or manufacturing operation, the practical draw is work that currently eats skilled hours. The tasks repeatedly named in the source material as targets for autonomous robots and drones are picking, put-away, goods inspection, line-side replenishment, moving heavy payloads and inspecting hazardous equipment (5) (4). One argument made directly is that robots with AI built in could gather information in hazardous environments without requiring workers to enter them (5). That is a safety win before you count a single minute saved.

Speed and reliability are the second argument. Autonomous systems, whether robots or drones, are described as changing how work is done in plants, power systems and mining sites. The stated aims are operations that are more efficient, safer for the people involved and more productive (5). Warehouse forklifts and autonomous mobile robots are already in commercial deployment: the German firm KION Group deploys self-driving forklifts and other autonomous mobile robots (3).

The third reason is staff, and it is the one staff will ask about. The guidance from industrial software chief technologist Arti Garg is that AI should augment rather than replace people in critical decision loops. Guardrails should decide where automated systems can act and where human supervisors remain responsible (5). Firms that frame it that way tend to get cooperation. Firms that do not tend to get refusal.

If your operation sits in manufacturing, our sector page sets out where this lands first.

How physical AI works, stage by stage

The mechanism is a loop, and it repeats continuously rather than running once from start to finish (4). It helps to walk through it stage by stage, because the cost and the risk in any project usually sit at a specific stage.

How it moves

One pallet pick, from camera frame to ledger row

A robot arm reads a label, plans the pick and stops for a person in the zone. The log is what you can audit.

Comes inA pallet reaches the inbound belt

1. 01

   Camera frame

   The camera captures the pallet. Mounts, lighting and a clean lens decide everything after this.

   Camera 2, inbound beltScreen

   Frame

   14:02:07, 1920 by 1080

   Pallet in view

   Ok: Yes, one pallet

   Lighting

   Ok: Even

   Lens

   Flagged: Smudge, lower left

2. 02

   Model reads the label

   Perception turns pixels into a fact: which pallet, which bin, which way round it sits.

   Label readScreen

   Pallet

   Ok: P-20418

   Bin

   A-14-03

   Orientation

   Turned 90 degrees

   Confidence

   Ok: 96 per cent

3. 03

   Arm plans the pick

   Grip, route and lift weight are worked out against the racking and the machine's limits.

   Pick plan, pallet P-204183 of 4

   • Done: Grip points chosen on the forks
   • Done: Route clear of racking, rows A and B
   • Done: Lift weight 18 kg, under the 25 kg limit
   • To do: Safety zone clear
4. 04

   Safety zone check

   If a person steps into the zone the arm stops. A supervisor decides when it may restart.

   A person decides here

   Safety zone 3Screen

   Arm

   Flagged: Stopped, person in the zone

   Detected

   Picker at bay A-13, 1.2 m away

   Restart

   Ok: Supervisor confirms the zone is clear

5. 05

   Logged as a ledger row

   Every pick, hold and failure becomes a row, so the floor can be audited like any other record.

   example-pick-log.xlsxSheet

   Pick log, one row per pick

   Matched: P-20418: held 38 s, then pickedMatched

   P-20419: picked

   Flagged: P-20420: label unreadableFlagged

Goes outThe pick is on the record

Gilt ring: a person makes the call at this step

Stage one: sensing. Cameras, lidar, radar, microphones and tactile or motion sensors collect information about the environment (4). This is the unglamorous layer: mounts, lighting, cabling, cleaning schedules for dirty lenses. Get the sensing wrong and no amount of clever software rescues the job.

Stage two: perception. Computer vision, machine vision, sensor fusion and simultaneous localisation and mapping may be used to identify objects, estimate their positions and build a representation of the surrounding world (4). In a warehouse this is the difference between “something is there” and “that is a Euro pallet, turned, half over the line”.

Stage three: planning. After interpreting its surroundings, the system uses task planning, motion planning or learned policies to select an action (4). This is where the machine decides which pallet to take, which route to run, which order to sequence, and how to get there without hitting racking.

Stage four: action. Control software converts the plan into commands for robot locomotion, motors, robotic joints or other actuators (4). New sensor information then lets the system evaluate the result and modify its plan as conditions change (4). The loop closes and starts again.

Stage five: running the loop across a fleet. Physical AI tasks have inference properties that differ sharply from digital AI. They consist of multiple rounds of inference and action execution, generating a chunk of actions in each inference round, and interleaving inference and execution asynchronously (1). Existing digital AI serving systems are unsuited to this pattern, which matters once you are running fleets of robots (1). The Kairos research system is designed to make the generate-execute loop a first-class citizen. It reduced average end-to-end task latency by 31.8 to 66.5 per cent over state-of-the-art digital AI serving practices, with gains scaling as the fleet grows (1).

Stage six: safety watching over the top. Because failures can cause physical damage or injury, these systems may require safety constraints, human oversight and fallback control mechanisms (4). One research proposal, written for hospital wards, puts a Runtime Safety Monitor at the centre. It evaluates multimodal physiological and system-level signals against defined constraints and targets three classes of event: deviations from the prescribed process, hardware and communication failures, and potential data tampering or misuse (2). On the commercial side, Nvidia’s Halos for Robotics, announced in June 2026, aims to make deployment safe for autonomous mobile robots in warehouses, humanoid robots inside factories and surgical robots (3).

A worked example: walking time at a Midlands parts distributor

The firm below is illustrative. It does not exist. The figures are our estimates to show the shape of a business case, and they carry no source markers.

Picture a 60-person distributor of industrial fasteners in the West Midlands, trading for about eighteen years, despatching roughly 900 order lines a day from a single unit. Three pickers walk the aisles full time. Two more staff spend most of their day on stock counts and rehandles because bin locations drift out of line with the system. Overtime runs most Fridays.

The physical AI step here is not a humanoid robot. It is a fleet of autonomous mobile robots moving totes between goods-in, the pick faces and the despatch bench. Floor sensors and a modest perception system guide them, and a person supervises the exceptions from a screen.

Setup: robots, charging points, a small fleet-serving computer, integration to the existing warehouse management system, floor marking changes and a safety assessment. Reasonable illustrative estimate: £180,000 to £240,000, with an eight to twelve month payback on labour and stock-count corrections alone. Walking time is the saving: pickers stop covering distance and start working at the face, and Friday overtime disappears in most scenarios.

Risks in this example are ordinary ones. A bin location that is wrong in the system becomes a physical stop. A goods-in process that changes shape mid-project forces rework. Two long-standing pickers dislike the change until they are given the exception-handling job and the pay that goes with it, after which they become the project’s best advocates.

The lesson transfers. Physical AI pays where movement is repetitive, routes are predictable, and the data underneath is accurate. It pays poorly where the floor changes every week or the system of record is fiction.

What it costs, and what goes wrong on a real floor

Costs fall into four buckets. Hardware (robots, sensors, compute). Integration (getting it to talk to your existing system of record). Data work (bin locations, layouts, product master data cleaned up). People (training, supervision, a named owner). Published UK bands for AI implementation work across 2026 are set out in our cost guide, and AGMM’s own fee structure is in what AGMM’s work costs.

Expect a safety line item that digital projects do not have. Industrial systems run where an unexpected decision can have consequences for safety, reliability and critical infrastructure (5). Hardware reflects this. Nvidia’s IGX Thor computing module for robotics and industrial applications carries an independent processor dedicated to safety-related workloads, on the argument that functional and safety systems need to run on the same silicon (3). Software side, the Halos operating system monitors every hardware block and every software library to spot failures quickly and isolates safety-critical computing workloads to avoid interference (3).

The failure modes are well documented. Challenges include incomplete sensor data, collision avoidance, real-time computing requirements, energy constraints and transferring behaviour learned in simulation to real machines (4). Safety definitions also differ by environment: a robotic vacuum cleaning a hallway and a robotic forklift handling heavy payloads at a loading dock raise completely different questions (3). And blind spots are real. As one Nvidia robotics lead put it, a robot coming down an aisle at six miles per hour cannot see what is around the corner. So robots tend to slow almost to a halt before a turn in unstructured factories, warehouses and hospitals (3).

Organisationally, two things go wrong. The first is automating a broken process: map it properly before you touch it, which is exactly what our process audit guide is for. The second is governance drift. Guardrails must decide where automated systems can act and where human supervisors remain responsible (5), and that decision needs writing down before the first machine moves.

What to do next if your floor is a candidate

Do not start with a robot. Start with the constraint. Map your highest-volume movement task on paper, time it properly, and ask whether the constraint is distance, waiting, rework or information (1) (5). If you have not automated anything yet, Five automations a UK business should test first is a gentler entry point than a capital project.

Second, score your readiness. Six questions about data quality, system of record and staff appetite will tell you more than a vendor demo: the six-question AI readiness assessment. Third, have the process looked at properly. Operations and delivery is one of the departments we work in, and the AI Constraint Audit is the first step: a fixed £2,000, 14-day audit that finds constraints, weak points and opportunities to save or earn serious money. A more extensive audit always follows and is included in the proposal.

Fourth, decide build or buy. Most firms are better served by automation of the surrounding software first, which is our business process automation service, before committing to physical kit. Fifth, bring your people in early rather than late; sceptical staff come around when the tool saves them an hour on the job they hate.

Every free resource we publish sits in AGMM’s free resources. For the wider cost picture before you commit budget, start with How much does AI implementation cost in the UK? (2026).

Frequently asked questions

What is physical AI, in plain English?

Physical AI means AI integrated into robots and machines so they can perceive and act in the physical world: they take in sensor data, work out what to do, and then do it (4). It covers industrial and warehouse robots, self-driving vehicles, drones and humanoid robots, and it is defined by the full loop of perceiving, planning and physically executing a task rather than by any single component (4). It overlaps with embodied AI and robotics, though the boundaries are not standardised (4).

Why do people say physical AI could change how businesses run?

Because it moves AI out of documents and into work that costs money: lifting, carrying, inspecting, counting and moving through hazardous areas (5). One industrial software executive cited an unnamed study suggesting industrial AI adoption rose by almost 78 per cent in two years (5). Nvidia’s physical AI business was described as already driving nearly $10 billion in annual revenue in March 2026 (3). The claim is not that everything changes at once. It is that tasks which previously needed a person walking a route can now be automated safely with human supervision (5).

What is a good physical AI example?

A warehouse of autonomous mobile robots carrying totes between goods-in and despatch, supervised from a screen, is the everyday example. Commercial deployments already include self-driving forklifts at the German KION Group and humanoid robots designed to work near people without isolated workstations, such as Agility Robotics’ Digit 5 (3). Nearer to home for most readers, robotic vacuum cleaners and lawn mowers are household physical AI operating with limited human control (4).

Where does physical AI help most in manufacturing?

Wherever movement or inspection is repetitive and routes are predictable: line-side replenishment, pallet movement, stock counting, and inspection of high-risk machinery (5). Robots can also collect information in hazardous environments so workers do not have to enter them (5). It helps least where the floor layout changes constantly or your stock data is unreliable, because perception and planning both depend on accurate data. Our manufacturing page in AI for manufacturing firms covers the sector-specific starting points.

What does physical AI cost a UK business?

More than software, less than a new wing on the building. Budget across hardware, integration, data cleanup and people, plus a real safety line for the monitoring and fallback mechanisms these systems require because failures can cause physical damage or injury (4). Published UK bands for AI implementation are set out in How much does AI implementation cost in the UK? (2026), and AGMM’s own charges are listed in what AGMM’s work costs. The cheapest first move is still an audit rather than a purchase.

Is investing in physical AI stocks a sensible move?

That is a different decision from adopting the technology, and shares are your call, not ours. What the source material shows is commercial scale: Nvidia has bet billions of dollars on physical AI and reported nearly $10 billion in annual revenue from the business in March 2026 (3), with safety systems such as Halos launched in 2025 and extended to robotics in June 2026 (3). For an operating business, the better question is which process on your floor is being held back, and the answer to that starts in the AI Constraint Audit.

Related guides

If this was useful, these go further on the same ground:

• What is RAG? Retrieval in plain English
• Five automations a UK business should test first
• AI and UK GDPR: what your business must get right
• How much does AI implementation cost in the UK? (2026)
• The plain-English AI glossary

Sources

Every numbered citation in this guide links to one of these. Each was fetched and read on 12 October 2026.

1. Kairos: A Scalable Serving System for Physical AI, arXiv.org.
2. Clinical Pathways as Safety Specifications for Physical AI in Hospital Wards, arXiv.org.
3. Nvidia’s big bet on physical AI aims for safer robotaxis, humanoid robots, Ars Technica.
4. Physical artificial intelligence - Wikipedia, en.wikipedia.org.
5. Building a safer path to autonomous industrial AI, MIT Technology Review.

How this guide was made: AGMM’s research engine gathered and read the sources above, a first model drafted the guide from them, a second model checked every sourced sentence against them, and an automatic rule check rejected any unsupported figure or broken link. It is marked for review by a named person at AGMM before it is treated as final. How we write and check our guides.