Physical AI in Industry: Current Landscape and Future Directions

Physical AI is having its moment. It is also about sixty years old.

What is physical AI?

IBM defines physical AI as AI systems that operate in and interact with the physical world. NVIDIA, which has done most to popularise the term, describes machines that perceive, understand, reason and act, and suggests that they previously could not perceive their surroundings.

That describes most factory robots, but not the field. Closing the loop between sensing, reasoning and acting has been robotics' central goal since the 1960s. What is new is how close AI has brought autonomous systems reliable enough for industry. So where do things stand in manufacturing, long one of the biggest users of robots?

Sixty-five years of structured certainty

The first industrial robot, Unimate, joined a General Motors line in 1961: a programmable arm moving between positions known in advance, running a fixed sequence. Most industrial robots still work this way. It has been hugely successful, but the economics are unforgiving: heavy engineering and capital committed to one configuration, much of it spent again when the product changes.

The alternative is almost as old. SRI's Shakey (1966–72) was the first mobile robot to reason about its own actions, and by the early 1980s GM's Consight system was using cameras to find randomly placed castings on a moving foundry conveyor. That is physical AI by today's definitions.

Two paths of innovation

Innovation in robotics follows one or both of two paths. The first creates new capability, automating tasks previously too difficult. The second cuts the engineering effort needed to deploy.

Vision took the first path: deep learning's breakthrough around 2012, with cheaper cameras and compute, made it dependable and mainstream. The collaborative robot took the second. When Danish plastics supplier Linatex bought a Universal Robots UR5 in 2008, it worked without a safety cage, but the real shift was that the company programmed it itself, with no prior experience. Until then, that needed specialist engineers.

Where the frontier sits

Repetitive jobs such as spot welding car bodies have been automated for decades. Beyond that, robots that see can pick parts from a jumbled bin or a moving belt. Robots that "feel" how hard they are pushing routinely fit parts together or grind surfaces, provided each part is held in a known place and every move is planned.

One of the biggest remaining challenges is dexterity: handling things as a skilled person does when nothing is quite where, or quite how, you expect. Parts in unknown positions, flexible materials such as cables, and longer jobs where the robot must recover when a step goes wrong. It is why so much assembly is still done by hand, and solving it would open up healthcare, agriculture, logistics and the home.

The current wave

AI that understands language can write robot programs and may let people instruct robots in plain words. AI that understands what it sees is learning to reason about a task: what comes next, and whether a step worked. AI is also becoming agentic, planning a job, calling on other tools and checking its own work. Google DeepMind presents Gemini Robotics ER 2, released in July 2026, as this kind of physical agent. NVIDIA's GR00T goes straight from camera images to robot movement. Affordable compute and accessible hardware underpin it all.

The reality check

None of this is plug-and-play. NVIDIA recommends at least 200 recorded demonstrations and a high-end GPU just to teach GR00T to move an apple onto a plate. By DeepMind's own figures, ER 2 misjudges how far a task has progressed almost half the time. Production lines need far better.

In January NVIDIA announced that "the ChatGPT moment for robotics is here"; in August, the chief executive of Chinese robot maker Unitree put it two to ten years away. Research, hype and deployment move at very different speeds.

Capability is only half the case. Each new model must still be adapted, integrated and validated, adding engineering effort rather than removing it. OEMs work on decade-long cycles, AI developers release monthly, safety standards change over years: a model superseded in months must run in a cell built to last fifteen years.

The UK picture

The UK has the world's top-ranked robotics research, but adoption lags. With 104 robots per 10,000 manufacturing workers, it is the only G7 economy below the global average, and 74% of UK manufacturing SMEs have no robot at all.

Yet momentum has never been stronger. New start-ups and spin-outs keep emerging, and UK robotics firms raised more in the first eight months of 2026 than in the whole of 2025. ARIA's Robot Dexterity programme targets exactly the bottleneck above. The Regulatory Innovation Office has made robotics a priority, alongside new Robotics Adoption Hubs. And in March 2027 the European Robotics Forum, Europe's leading robotics gathering, comes to Birmingham, putting the UK at the centre of European robotics.

The path to deployment is already mapped. techUK's Seizing the robotics opportunity report calls for a Robotics Opportunities Action Plan, building on the MTC's Global Robotics Clusters Study. Both point to the same steps: join up existing efforts; back regional clusters with vendor-neutral testbeds where firms can prove capability before buying; offer tiered adoption support for SMEs; use public procurement to create early customers; and invest in the skills and standards these systems need.

A reason for optimism

Physical AI is not new, but it has never moved this fast or attracted this much capital, and the goals set in the 1960s have never looked closer.

For the UK, the prize may be bigger than for most. Much of British manufacturing is high-value, low-volume work: aerospace, defence, pharmaceuticals, medical devices, and thousands of SMEs making many variants in small batches. Traditional automation pays back only when the same part is made over and over, which is part of the reason UK robot density has stayed so low. Robots that cope with variation and need far less engineering change that equation. If physical AI delivers on both paths, it won't replace the robots that already serve industry so well. It will open up applications out of reach for decades, and few countries stand to gain more than the UK.