Why Intelligence Does Not Easily Transfer to Machines

Artificial intelligence has made remarkable progress in the last decade. Models can understand images, generate language, write software, recognise patterns and increasingly reason across different forms of information. The natural next step is to put this intelligence into machines that can interact with the physical world. This has become the central idea behind physical AI and the current generation of humanoid robotics. The proposition is straightforward: give a robot enough sensors, enough data and a sufficiently capable model, and it should eventually learn how to operate in the real world. But there is a fundamental difference between processing information and operating in reality. A model can produce an incorrect sentence and nothing physical happens. A robot can make the same kind of mistake and move an object, apply force, damage equipment or interact with a person incorrectly. This makes robotics fundamentally different from most digital AI applications. The problem is not simply putting a better model inside a robot. The problem is making that intelligence reliable when it is connected to a physical system.

A large part of modern AI is built around learning from previous examples and using those learned relationships to make predictions about new situations. This is extremely powerful, but the physical world does not repeat itself in a controlled manner. A robot can be trained on thousands of examples of grasping an object and still encounter a slightly different object, surface, weight, position or environmental condition that changes the correct action. The model still has to make a decision even though it has never encountered that exact situation. This is one of the major challenges in general-purpose robotics. Research in the field continues to identify generalisation and changes between training environments and real-world environments as major limitations. The answer is not simply to collect more data. Physical data is expensive to generate, because a robot has to physically perform an action to produce an example. More importantly, it is impossible to collect examples of every combination of objects, environments, people and physical conditions that a robot might eventually encounter. Data describes what has happened before, while a robot has to operate in what is happening now.

This creates a deeper problem with using probabilistic AI for physical action. Modern models can produce highly capable outputs while still being uncertain or incorrect. In a digital system, an incorrect prediction can often be reviewed, rejected or generated again. A robot eventually has to turn its prediction into movement. If it incorrectly identifies an object, estimates its position incorrectly or chooses an inappropriate action, the error becomes physical. This means that a robot cannot simply depend on a model that is correct most of the time. It needs to recognise when its prediction may be unreliable and continuously compare its expectations with what is actually happening. A robot needs to observe the result of its actions and adjust its behaviour accordingly. This is where traditional feedback control remains important even as AI models become more capable. Intelligence in a physical machine cannot stop at prediction. It has to include the ability to detect errors, respond to them and recover from them. A general-purpose robot will inevitably make mistakes, so the ability to recover from an incorrect assumption may be as important as making the original prediction correctly.

The real world also contains considerably more context than a robot can directly observe. A camera can show a robot that a box is present, but it does not tell the robot how heavy the box is. A robot may see a person walking toward it without knowing whether that person intends to stop, turn or continue. It can recognise an object without knowing whether moving it is safe or whether it is important to the task. Humans constantly use information that is not explicitly represented in their immediate visual input. We use previous events, physical experience, environmental cues and continuous feedback to understand what is happening. Robots are still developing this ability. Current vision-language-action systems are making significant progress by combining visual information, language and action, but the fundamental problem remains: the machine only has access to a limited representation of reality and has to make decisions from that representation. The challenge becomes even harder during long tasks, where the robot must remember what has happened, understand how the current state resulted from previous actions and determine whether its assumptions are still valid. A system may perform several steps correctly and then fail because one small change in the environment invalidates the assumptions behind the original plan.

Simulation and large-scale training provide another partial solution but introduce their own limitations. Simulation allows robots to perform enormous numbers of experiments without physically damaging hardware, making it extremely valuable for learning movement, navigation and interaction. However, a simulated environment is still an approximation of reality. Materials, friction, contact forces, sensor noise, mechanical tolerances and environmental variation are difficult to reproduce perfectly. A robot can therefore learn a behaviour that works extremely well in simulation and still encounter unexpected behaviour when transferred to the physical world. This is not an argument against simulation. It means that simulation should be treated as one source of experience rather than a complete representation of reality. The robot ultimately has to learn from physical interaction as well. This creates a difficult data problem because real-world experience is slow and expensive compared with digital data. If a robot needs to learn how to recover from thousands of different failures, those failures have to occur somewhere. The machine has to experience them, measure them and learn from them. Scaling physical experience is therefore fundamentally different from scaling text or image datasets.

There is another constraint that becomes particularly important when AI moves from data centres into machines: computation. The most capable AI models require substantial memory, processing power and energy. A data centre can provide large amounts of computing capacity and cooling, but a robot has a finite battery, limited thermal capacity, limited physical space and a hardware cost that directly affects deployment. Running a capable model locally gives the robot greater independence and can reduce communication latency, but requires powerful onboard hardware. Sending inference to the cloud provides access to much larger computing resources but introduces network dependency and communication latency. For some tasks this may be acceptable, but a robot interacting continuously with its environment cannot always wait for a remote system to process every observation. This creates a difficult balance between model capability, latency, power consumption, hardware cost and reliability. Smaller models that can run locally are becoming increasingly important for this reason, but reducing model size also creates a tradeoff between computational efficiency and capability. The challenge is therefore not simply to build the largest possible model. It is to build an intelligence system that provides sufficient capability within the physical limits of the machine.

This also exposes a broader problem with the way robotic intelligence is often discussed. The intelligence of a robot cannot be separated completely from its body. Its sensors determine what information is available. Its mechanical structure determines what it can physically do. Its actuators determine how accurately it can execute an action. Its battery determines how long it can operate. Its computing hardware determines how quickly it can process information. A powerful model attached to an inadequate physical system does not create a powerful robot, just as a highly capable mechanical platform with an unreliable decision system does not create a reliable robot. The complete system has to work together. This becomes especially important for humanoids because the human environment is not designed around robots. Floors are uneven, objects are placed at arbitrary positions, tools are designed around human hands, people move unpredictably and objects can be damaged, misplaced or partially obstructed. A robot operating in this environment has to deal with variation continuously. Humans handle much of this without explicit calculation. We adjust our movements based on what we see and feel, notice when an object is heavier than expected and change our approach when an action does not work. Robots are still developing this form of continuous adaptation.

This is where the distinction between artificial intelligence and physical intelligence becomes important. Artificial intelligence can be extremely capable at processing representations of the world, but physical intelligence requires a machine to maintain a reliable relationship with the world itself. The machine has to act, observe the result, compare reality with its expectations and modify its behaviour. It has to learn from previous experience without becoming dependent on previous situations. It has to use what it has learned while remaining capable of adapting when the current environment does not match its expectations. It also needs some understanding of uncertainty. A robot should not treat a situation that closely matches its experience in the same way as one where its assumptions are weak. This is one of the areas where future robotic systems may need to move beyond simply scaling existing AI approaches. More parameters, more data and more computation can improve capability, but none of them independently solves the problem of grounding that capability in the physical world.

The next generation of robotics will therefore not be defined only by who builds the largest model or collects the largest dataset. The harder problem is building systems that continue to function when the model is uncertain, the environment changes, information is incomplete and the physical system behaves differently from what was expected. Physical AI is not simply another application of artificial intelligence. It is a systems problem that combines learning, perception, control, computation, mechanical design and real-world interaction under constraints that do not exist in purely digital environments. The progress of foundation models has brought robots significantly closer to general-purpose behaviour, but the remaining challenge is not only making machines more intelligent. It is making that intelligence dependable when it meets reality. A robot does not operate in a dataset. It operates in the world, and the world does not provide clean boundaries, complete information or predictable conditions. The real challenge for physical AI is therefore not to make robots that can predict what usually happens. It is to build machines that can continue to function when what happens is something they have never seen before.