When Machines Learn to Act in the World

Robots have been moving, sensing and responding to their surroundings for decades. What changes when machines begin to learn physical skills that can be applied beyond the tasks for which they were designed?

Robots have become difficult to ignore. We see humanoid machines walking through factories, robotic arms handling unfamiliar objects, and increasingly convincing demonstrations of machines navigating the physical world. Alongside these developments, a term keeps appearing: Physical AI.

It sounds like the next great technological breakthrough. But what exactly is new about it?

Robots have existed for decades. Industrial machines assemble cars, move goods and perform intricate operations with extraordinary precision. They alreadyi have motors, sensors, cameras and computers. Some have used artificial intelligence for years.

The interesting distinction is not between machines with and without intelligence. It is between machines that become increasingly capable at specialised tasks and machines that can acquire physical skills applicable across different situations.

A robot that learns to handle one particular object is useful. A robot that can apply what it has learned to unfamiliar objects, environments and tasks represents a considerably more ambitious possibility.

That is the promise behind much of today's Physical AI research. And to understand why it matters, it helps to look at how machines acquired their capabilities in the first place.

The long history of the machine

Human beings have always invented ways to extend their physical capabilities. A lever allows us to lift something too heavy for our muscles. A wheel makes it easier to transport a load. Gears transmit and transform movement.

These early mechanisms extended what our bodies could accomplish, but people remained responsible for directing their operation.

Over time, machines became more elaborate. Waterwheels and windmills harnessed natural forces. Clockwork mechanisms produced carefully regulated movements. Steam engines and, later, electric motors made mechanical power available on an unprecedented scale.

The Industrial Revolution transformed the relationship between people and machines. Mechanical power could now perform work continuously, at speeds and scales that human muscles could never sustain.

Yet a machine could be extraordinarily powerful without knowing anything about the work it performed. A spinning machine did not understand thread. A steam engine did not understand the factory it powered.

Something else was needed to make machines responsive to their surroundings.

One of the most consequential developments was the feedback mechanism.

Consider a thermostat. It measures the temperature, compares it with a desired value and adjusts the heating accordingly. The system uses information about its environment to regulate its own operation.

This principle is much older than modern computing. Mechanical governors regulated the speed of steam engines, while earlier control mechanisms can be found throughout engineering history.

In the twentieth century, these ideas became central to cybernetics, the study of control and communication in animals and machines.

Feedback introduced a different relationship between a machine and its environment. Rather than simply executing a movement, a machine could respond to the consequences of its own actions.

Electronics and computers expanded these possibilities enormously. Machines could process sensor data, execute complex sequences and adjust their operations through software.

Industrial robots brought these developments together. They could weld, paint, assemble and manipulate objects with remarkable accuracy.

But much of this success depended on carefully designed environments. The robot knew where an object would appear, how it was positioned and which movement was required. Engineers had anticipated the conditions under which the machine would operate.

The history is not a simple progression from unintelligent to intelligent machines. Mechanical power, feedback, computation and learning have developed alongside one another, gradually expanding what machines can accomplish.

Physical AI builds upon that history. Its ambition is to make machines less dependent on having every relevant situation anticipated in advance.

What makes Physical AI different?

Physical AI describes the application of artificial intelligence to systems that perceive, interact with and act upon the physical environment.

This includes robots, autonomous vehicles and other machines whose decisions have physical consequences.

The concept itself is not entirely new. Robotics has employed computer vision, machine learning and intelligent planning for decades. Researchers have long explored how machines might learn movements, recognise objects and adapt to changing circumstances.

What has changed is the sophistication of the AI models available, the scale at which they can be trained and the growing possibility of transferring learned capabilities between tasks.

Traditional automation relies heavily on engineers specifying what a machine should do and under which circumstances. Learning-based approaches allow some capabilities to be acquired from data and experience.

A robot might learn to recognise objects, estimate their positions, select ways of grasping them and adjust its movements when something unexpected happens.

Through reinforcement learning, machines can develop behaviours by attempting tasks and learning from their outcomes. Simulation allows many of these experiments to take place in virtual environments before the resulting behaviours are tested in reality.

More recently, advances in foundation models have introduced another possibility: connecting visual perception, language and physical action.

Imagine instructing a robot to put a red cup in a cupboard.

The robot must identify the cup, understand the instruction, locate the cupboard, plan a sequence of movements and carry them out. If the cup is positioned differently than expected, it must adjust.

Each of these capabilities has been studied independently. What makes their combination interesting is the possibility of developing systems that can apply them across many different situations.

Researchers are exploring models that learn relationships between what a machine sees, what it is asked to do and the physical actions required.

The ambition is not merely to produce a robot that performs one task more efficiently. It is to develop capabilities that can be reused and combined when the task or environment changes.

But there is an important difference between a model that can interpret many instructions and a machine that can reliably execute them.

That difference becomes clear when intelligence encounters the physical world.

Intelligence meets the physical world

There is a striking contrast between the recent success of artificial intelligence in digital environments and its progress in the physical world.

Large language models can generate text, translate languages, write software and reason through complex problems. They operate primarily through representations of information: words, images, code and other forms of data.

A physical robot faces a rather different challenge.

A cup has weight. A surface has friction. An object can slip, bend or break. Gravity acts continuously, and movements have momentum.

A language model can produce a convincing description of picking up a glass. A robot must actually pick it up.

And if it gets the movement wrong, the glass may fall.

This is more than an additional engineering difficulty. Physical intelligence operates under conditions that are fundamentally different from generating information.

A robot must coordinate perception, planning and movement while continuously responding to an environment that cannot be completely predicted.

Recognising a glass is not the same as knowing how firmly to grasp it. Planning a movement is not the same as executing it while maintaining balance. Understanding an instruction does not guarantee that the machine possesses the physical capabilities required to carry it out.

Some of this knowledge can be represented through explicit physical models. Other aspects may be learned through experience.

Modern AI offers increasingly powerful approaches to that learning process. But it does not eliminate the need for mechanical engineering, control theory or an understanding of physical constraints.

A sophisticated robot remains a combination of motors, actuators, sensors, processors and software. Many essential functions, including stability, positioning and basic motion control, may still depend on conventional engineering rather than generative AI.

Physical AI does not replace the machine. It may expand what the machine can learn to do.

Why this moment is different

If robotics, feedback and machine learning have existed for decades, why has Physical AI become such a prominent subject now?

Part of the answer lies in the convergence of several technological developments.

Modern AI systems can learn increasingly sophisticated representations from large quantities of data. Computer vision has improved dramatically. Processing hardware has become more capable, while simulation environments allow robots to practise tasks without requiring every experiment to take place in the physical world.

Developments in motors, actuators, batteries and sensors also contribute to what robots can achieve.

Perhaps most importantly, researchers are attempting to apply some of the principles behind foundation models to physical action.

The success of large models in language and vision raises an intriguing possibility. Could models trained across many tasks and environments acquire more general physical capabilities?

There are reasons for optimism, but also important differences.

The internet provides enormous quantities of text and images for training AI systems. Comparable datasets of physical interactions are much harder to collect.

A robot must experience not only what the world looks like, but what happens when it acts upon it.

How does a particular object respond to pressure? What happens when a surface is slippery? How much force is needed to open an unfamiliar door?

These interactions involve physical properties that cannot always be inferred reliably from images or descriptions.

Simulation helps address this problem. Robots can practise movements, experience failures and improve their behaviour in virtual environments. But simulations never reproduce the real world perfectly.

A movement that succeeds in simulation may fail when confronted with unexpected friction, different materials or small variations in an object's position.

The challenge is therefore not simply to develop a more powerful AI model. It is to connect learning, perception and physical control in ways that remain dependable outside the conditions in which the system was trained.

This is where the ambition of general-purpose robotics becomes particularly demanding.

From specialised machines to transferable skills

Industrial robots are already exceptionally capable within their intended environments. A robotic arm may perform thousands of operations with precision and consistency.

But a robot designed to assemble a particular component is not necessarily capable of tidying a kitchen, opening an unfamiliar door or helping someone who has dropped an object.

These tasks demand flexibility.

Human beings possess an enormous amount of practical knowledge about the physical world. We understand, often without conscious thought, how objects behave, how to maintain our balance and how to adjust a movement when something goes wrong.

Much of this knowledge is difficult to express as explicit instructions.

Learning-based robotics offers a possible route towards acquiring some of these capabilities through experience rather than specifying every behaviour individually.

A machine might learn a collection of physical skills and apply them in combinations that were not separately programmed. It could interpret instructions, select appropriate actions and adapt its movements to the circumstances.

That would represent a significant departure from the way many industrial robots are deployed today.

Yet the distinction between a convincing demonstration and a dependable machine remains substantial.

A robot may successfully perform a task in a laboratory while struggling with a seemingly minor variation. It may recognise an unfamiliar object but fail to manipulate it. It may complete an operation repeatedly before encountering a situation that its training did not adequately prepare it for.

General-purpose capability is not established by demonstrating many individual tasks. It depends on how reliably a machine can transfer what it has learned to situations it has not previously encountered.

And even that is only part of the question.

A general-purpose robot must also be useful.

A specialised machine may remain faster, cheaper and more dependable at a particular task than a flexible robot capable of doing many things.

The economic significance of Physical AI may therefore emerge in places where conventional automation has been difficult: environments that change frequently, tasks that vary considerably or situations in which programming every possible action would be impractical.

Warehouses, construction sites, hospitals and domestic environments present very different challenges, but they share a degree of physical variability that makes automation difficult.

The question is not merely whether a robot can perform a task. It is whether it can do so consistently, safely and economically enough to justify its deployment.

That is where advances in physical intelligence could begin to matter beyond research laboratories.

A new era of machines?

Looking back across the history of machines, Physical AI appears less like a sudden invention than the convergence of developments that have been unfolding for centuries.

Mechanical power extended human strength. Feedback made machines responsive to their surroundings. Computing made increasingly complex behaviour programmable. Robotics combined these capabilities to perform physical tasks with considerable autonomy.

Learning-based robotics adds the possibility of acquiring behaviours that are not individually specified in advance.

In that sense, the development is evolutionary. It builds upon engineering disciplines and ideas that have been developing for decades.

But technological history also shows that evolutionary improvements can eventually produce much larger changes.

The electric motor was not simply a better source of mechanical power. Its flexibility helped transform how factories could be organised, allowing machinery to be distributed and arranged differently.

Computing was not merely a more sophisticated form of automation. It created new possibilities for processing information and coordinating activities across organisations and societies.

Physical AI raises the possibility of a comparable change in the relationship between machines and physical work.

If machines become capable of learning transferable physical skills, they may eventually operate in environments that have resisted automation. They could move beyond carefully structured production lines into settings where tasks, objects and circumstances change continuously.

That would affect not only what machines can do, but where they can be used and how people organise work around them.

We are not there yet.

The foundations are real, and progress in learning-based robotics is substantial. But dependable, economically useful performance across unfamiliar physical environments remains an open challenge.

The important question is therefore not whether machines have suddenly become intelligent. Nor is it whether Physical AI represents an entirely new branch of technology.

It is whether the capabilities now emerging can make the transition from specialised demonstrations to reliable, transferable physical skills.

For centuries, we have built machines to extend what our bodies can do. We are now exploring how far they can extend our capacity to perceive, learn and act in the physical world.

Whether that constitutes a new era will become clear not from what robots demonstrate in laboratories, but from what they can reliably accomplish in the world beyond them.