No LLM generated, assisted/co-written, or edited work.
Read full explanation
TL;DR: Unlike machines, human beings discover that they have a body by exploring and moving, not by following instructions. The text raises the question of whether a body is necessary for AI learning, not merely as a channel for data, but as a bridge to firsthand experience with the real world.
Imagine a baby just a few weeks into life, cozy in his crib, looking up at a mobile hanging above his head while doing the things babies do.
With tightly clenched fists, the baby moves its arms and hands energetically in all directions, as well as its legs, with no apparent control or at least without any real awareness that those hands, arms, and legs are part of its physical body and that it is the one actually moving them.
No one has told the baby that those limbs are theirs and that they can do incredible things with them like running, playing soccer, or painting, and doing so would make no sense, since this baby is just beginning to discover the world and would find it difficult, if not impossible, to understand: “Hey, those are your hands. Use them this way!” Furthermore, having been born so recently, the baby won’t understand spoken language on a semantic level, much less written language, and will hardly be able to distinguish where its body ends and the outside world begins.
So, what will actually happen is that this erratic limb behavior will continue over the following weeks until around 4 months of age, when a gradual learning process will begin in which the baby starts to discover on its own that it can control its limbs through hand-eye coordination, a behavior known as “hand regard”.
This observation process involves both physical self-exploration and complex neural development; that is, the baby begins to identify empirically and through observation based on their movements and cognitive processes what is part of their body and what is not.
In an agnostic manner, this gradual process of associating a certain type of signal understood as an intention to act with the search for an observable change in the environment that determines that it is that signal (my signal), and not an external agent, that actually generates that change, is what we might call, in general terms, the discovery of the body.
From a biological perspective for human beings, this process of progressive learning is quite organic and even transparent; that is, we do not realize it is happening until, at some point, we become aware of our body and our control over it.
In this context, the question arises: Can this principle of discovering the body through experimentation and observation also be applied to machines? Or rather,
Can an AI or some form of computation we might classify as an “AI Mind” discover its own physical body based on the same principle of experimentation and observation, without any pre-established configuration or programming for that specific body (hardware) or any other?
The next challenge
Giving an AI or a similar computing system a body does not seem to be the greatest challenge in modern robotics. Humanoid robots like Tesla’s Optimus Prime, or Boston Dynamics’ robot dog “Spot,” are capable of performing increasingly complex movements, ranging from acrobatics and kung fu sequences to artificially replicating the movements of the human hand.
However, to achieve this level of evolution, a prototype of the body with minimal initial hardware that is sufficiently viable both economically and functionally, while the cognitive layer, the AI mind that will control that body, is developed using a deep reinforcement learning (DRL) model in a simulated environment with a virtual representation of the robot, where it learns, practices, and refines its movements. Paradoxically, this learning takes place in a simulated body, a digital representation of the physical hardware, meaning the robot learns to move without a real body (disembodied AI). The transfer of this simulated learning to the real world is known as the “sim-to-real gap”, and it represents one of the most significant challenges in embodied AI, because simulators are not a mathematically perfect or faithful representation of the real world.
In this context, even though the AI mind is powerful enough to recognize a chair without any major problems, it has never had the physical, real-world experience of sitting in one. Once again, the firsthand experience provided by the body is missing.
This strategy of coupling a physical body with an AI mind trained without one addresses the challenge of learning from a classical top-down perspective, that is, the body first and the mind second, creating a real gap (sim-to-real gap) when the learning is transferred to the body.
What tends to happen in practice is that the body comes before the mind or at least a few steps ahead of it. For example, let’s consider a production line at an auto assembly plant consisting of multiple stations. Some of these stations are fully automated by robotic arms that must support the weight of a complete car chassis and also perform certain movements with it in the air. The robotic arm is designed and built specifically for that purpose, to support weight and perform certain movements in a specific way. In other words, the programming, if it didn’t happen afterward, occurred at roughly the same pace, but it is programming specific to a specific body.
The point is that, from this perspective, where the body comes before the mind, software design will always be carried out with a specific goal in mind: From an assembly line, to playing chess…
Of course, there is no “discovery of a body” in playing chess. This robotic arm is simply following preprogrammed instructions or steps designed to achieve the objective for which it was programmed; the same applies to the case of a robotic arm on an assembly line.
In this scenario, proposals have been put forward that involve a different paradigm shift for the learning process; rather than creating an AI mind with complex models encapsulated in a virtual environment (Top-Down Model), it is better to develop systems structured in layers of that directly connect perception with action, using the real world as its own model (Bottom-Up Model). This new bottom-up approach, based on direct and continuous physical interaction with the environment, seeks to imitate nature’s biological evolution rather than relying on preprogrammed abstract representations, and it represents the fundamental basis for the discovery of the body, as it allows for:
Adaptability to new morphologies. An Artificial General Intelligence (AGI) could operate in different physical configurations, that is, in different bodies, without the need for pre-training or models that tell it who it is or what it can do; rather, it could adapt autonomously through experimentation.
Active learning through causal inference. Through random experiments, AGI can generate actions in the form of signals and observe how the states of objects around it change. Causal inference uses the action-perception mechanism to determine what is my body and what is not.
Self-cognition and the mirror test. Similar to how babies develop control over their bodies through eye-hand or hand-regard experimentation, an AGI could not only arrive at the same discoveries, but also, in a physical environment where other objects exist, could come to identify itself and develop self-awareness through something called the mirror test.
Does the body matter for thinking, or is it just a channel for inputting data?
The Divided Mind
For a long time, studies suggested something like this: the mind is not a unified whole.
There is a barrier between what we perceive and what we think.
On one hand, there is perception. It is automatic, almost “blind”: it receives information from the world and passes it on to thought, without further thought.
On the other hand, there is thought. That does process everything thoroughly, without so many limitations.
In other words: one system receives data from the environment. The other turns it into complex ideas.
The Flaw in the Theory
But there’s a problem with this. If it were entirely true, it wouldn’t matter whether we saw the world with our eyes or “saw” it with sonar. The body would be nothing more than a data input cable. Nothing else.
And that’s when other researchers began to say: Wait, not so fast.
The Body Takes Center Stage
The argument that began to gain traction was that the body does matter. That seeing, moving, and touching, all of that is part of how we learn; it’s not just a conduit.
Think of an apple. When you look at it, you don’t first think about the exact shade of red, or how round it is, or its texture. The first thing you see is the action: you pick it up and eat it.
This was called “affordances”, what the environment offers you the opportunity to do. They aren’t just properties of the object: they are opportunities for action.
Three Reasons Not to Separate Them
Perception is already knowing what to do. Seeing your surroundings isn’t just about seeing shapes or colors. It’s about seeing how to move within them and what you can do with things.
Perception controls action, and vice versa. Moving, using your hands, none of that is automatic or a blind command from the brain. It’s a constant back-and-forth between what you see and what you do.
The object and the body go together. You can throw a ball because it’s light, round, and easy to grab. The object and your body fit together like two pieces of the same system.
From Theory to Machine
This idea didn’t stay confined to the books.
Some research goes even further:
If this works to explain how humans learn, it could also serve as a guide for designing a physical AI?
This post is an original piece written for educational purposes, extracted and adapted directly from my master’s thesis in Computer Engineering, with a specialization in Artificial Intelligence, of which I am the author.
TL;DR: Unlike machines, human beings discover that they have a body by exploring and moving, not by following instructions. The text raises the question of whether a body is necessary for AI learning, not merely as a channel for data, but as a bridge to firsthand experience with the real world.
Imagine a baby just a few weeks into life, cozy in his crib, looking up at a mobile hanging above his head while doing the things babies do.
With tightly clenched fists, the baby moves its arms and hands energetically in all directions, as well as its legs, with no apparent control or at least without any real awareness that those hands, arms, and legs are part of its physical body and that it is the one actually moving them.
No one has told the baby that those limbs are theirs and that they can do incredible things with them like running, playing soccer, or painting, and doing so would make no sense, since this baby is just beginning to discover the world and would find it difficult, if not impossible, to understand: “Hey, those are your hands. Use them this way!” Furthermore, having been born so recently, the baby won’t understand spoken language on a semantic level, much less written language, and will hardly be able to distinguish where its body ends and the outside world begins.
So, what will actually happen is that this erratic limb behavior will continue over the following weeks until around 4 months of age, when a gradual learning process will begin in which the baby starts to discover on its own that it can control its limbs through hand-eye coordination, a behavior known as “hand regard”.
This observation process involves both physical self-exploration and complex neural development; that is, the baby begins to identify empirically and through observation based on their movements and cognitive processes what is part of their body and what is not.
In an agnostic manner, this gradual process of associating a certain type of signal understood as an intention to act with the search for an observable change in the environment that determines that it is that signal (my signal), and not an external agent, that actually generates that change, is what we might call, in general terms, the discovery of the body.
From a biological perspective for human beings, this process of progressive learning is quite organic and even transparent; that is, we do not realize it is happening until, at some point, we become aware of our body and our control over it.
In this context, the question arises: Can this principle of discovering the body through experimentation and observation also be applied to machines? Or rather,
The next challenge
Giving an AI or a similar computing system a body does not seem to be the greatest challenge in modern robotics. Humanoid robots like Tesla’s Optimus Prime, or Boston Dynamics’ robot dog “Spot,” are capable of performing increasingly complex movements, ranging from acrobatics and kung fu sequences to artificially replicating the movements of the human hand.
However, to achieve this level of evolution, a prototype of the body with minimal initial hardware that is sufficiently viable both economically and functionally, while the cognitive layer, the AI mind that will control that body, is developed using a deep reinforcement learning (DRL) model in a simulated environment with a virtual representation of the robot, where it learns, practices, and refines its movements. Paradoxically, this learning takes place in a simulated body, a digital representation of the physical hardware, meaning the robot learns to move without a real body (disembodied AI). The transfer of this simulated learning to the real world is known as the “sim-to-real gap”, and it represents one of the most significant challenges in embodied AI, because simulators are not a mathematically perfect or faithful representation of the real world.
In this context, even though the AI mind is powerful enough to recognize a chair without any major problems, it has never had the physical, real-world experience of sitting in one. Once again, the firsthand experience provided by the body is missing.
This strategy of coupling a physical body with an AI mind trained without one addresses the challenge of learning from a classical top-down perspective, that is, the body first and the mind second, creating a real gap (sim-to-real gap) when the learning is transferred to the body.
What tends to happen in practice is that the body comes before the mind or at least a few steps ahead of it. For example, let’s consider a production line at an auto assembly plant consisting of multiple stations. Some of these stations are fully automated by robotic arms that must support the weight of a complete car chassis and also perform certain movements with it in the air. The robotic arm is designed and built specifically for that purpose, to support weight and perform certain movements in a specific way. In other words, the programming, if it didn’t happen afterward, occurred at roughly the same pace, but it is programming specific to a specific body.
The point is that, from this perspective, where the body comes before the mind, software design will always be carried out with a specific goal in mind: From an assembly line, to playing chess…
Of course, there is no “discovery of a body” in playing chess. This robotic arm is simply following preprogrammed instructions or steps designed to achieve the objective for which it was programmed; the same applies to the case of a robotic arm on an assembly line.
In this scenario, proposals have been put forward that involve a different paradigm shift for the learning process; rather than creating an AI mind with complex models encapsulated in a virtual environment (Top-Down Model), it is better to develop systems structured in layers of that directly connect perception with action, using the real world as its own model (Bottom-Up Model). This new bottom-up approach, based on direct and continuous physical interaction with the environment, seeks to imitate nature’s biological evolution rather than relying on preprogrammed abstract representations, and it represents the fundamental basis for the discovery of the body, as it allows for:
Adaptability to new morphologies. An Artificial General Intelligence (AGI) could operate in different physical configurations, that is, in different bodies, without the need for pre-training or models that tell it who it is or what it can do; rather, it could adapt autonomously through experimentation.
Active learning through causal inference. Through random experiments, AGI can generate actions in the form of signals and observe how the states of objects around it change. Causal inference uses the action-perception mechanism to determine what is my body and what is not.
Self-cognition and the mirror test. Similar to how babies develop control over their bodies through eye-hand or hand-regard experimentation, an AGI could not only arrive at the same discoveries, but also, in a physical environment where other objects exist, could come to identify itself and develop self-awareness through something called the mirror test.
The Divided Mind
For a long time, studies suggested something like this: the mind is not a unified whole.
There is a barrier between what we perceive and what we think.
On one hand, there is perception. It is automatic, almost “blind”: it receives information from the world and passes it on to thought, without further thought.
On the other hand, there is thought. That does process everything thoroughly, without so many limitations.
In other words: one system receives data from the environment. The other turns it into complex ideas.
The Flaw in the Theory
But there’s a problem with this. If it were entirely true, it wouldn’t matter whether we saw the world with our eyes or “saw” it with sonar. The body would be nothing more than a data input cable. Nothing else.
And that’s when other researchers began to say: Wait, not so fast.
The Body Takes Center Stage
The argument that began to gain traction was that the body does matter. That seeing, moving, and touching, all of that is part of how we learn; it’s not just a conduit.
Think of an apple. When you look at it, you don’t first think about the exact shade of red, or how round it is, or its texture. The first thing you see is the action: you pick it up and eat it.
This was called “affordances”, what the environment offers you the opportunity to do. They aren’t just properties of the object: they are opportunities for action.
Three Reasons Not to Separate Them
Perception is already knowing what to do. Seeing your surroundings isn’t just about seeing shapes or colors. It’s about seeing how to move within them and what you can do with things.
Perception controls action, and vice versa. Moving, using your hands, none of that is automatic or a blind command from the brain. It’s a constant back-and-forth between what you see and what you do.
The object and the body go together. You can throw a ball because it’s light, round, and easy to grab. The object and your body fit together like two pieces of the same system.
From Theory to Machine
This idea didn’t stay confined to the books.
Some research goes even further:
This post is an original piece written for educational purposes, extracted and adapted directly from my master’s thesis in Computer Engineering, with a specialization in Artificial Intelligence, of which I am the author.
Image from: Pavlina Baudysova
I am still working on it, but in the meantime, let me know your comments, questions, and thoughts!