Juan Manuel Lucero
11/2025

Brains in jars, grounding, or why AI needs a body

Large language models are disembodied brains floating in digital space…

Someone who has read every cookbook in the world knows in theory how to make a tortilla, but if they have never cracked an egg, they have no idea how much force it takes to do it without making a mess. They can define the word "near," but they have no physical intuition of how far near is.

That is, in essence, the great drama of today's artificial intelligence. Large language models are disembodied brains floating in digital space. They can write brilliant essays on phenomenology or recite the maintenance manual of a nuclear plant, but they do not know what it feels like to hold a tool, and they do not understand the resistance of materials. And that is the limit: if scientists want AI to be truly useful in the real world, in factories, hospitals or homes, computing power is not enough. It needs a body.

This disconnect is not a minor problem. It is what cognitive science calls the "symbol grounding" problem. AI knows that the word "apple" is related to "fruit," but it has no sensory experience that anchors that concept to physical reality.

The idea has deep roots. In 1991, Rodney Brooks at MIT published a foundational paper: Intelligence without Representation. Brooks argued against the dominant Cartesian paradigm, which saw the mind as a logical processor separate from the body. He held that intelligence does not emerge from building complex, abstract mental maps but from direct, messy, physical interaction with the environment (perhaps influenced by Dreyfus or Merleau-Ponty: intelligence is "being in the world"?). The world, Brooks said, is its own best model. But for decades we lacked the technology to join that philosophical vision with modern processing power. Robots could move, but they did not understand what they were moving.

Today that gap is closing thanks to a new technical architecture: we are watching the birth of what recent papers (such as the survey by Ma et al., 2024) classify as Vision-Language-Action models (VLAs). The great innovation here is not mechanical but semantic. Models such as RT-1 and RT-2 showed something remarkable: if we treat a robot's actions (move arm, open gripper) as if they were "words," or tokens, we can use the same Transformer architecture behind ChatGPT to control a body. That allows something unprecedented, which is to transfer the vast knowledge of the internet to the physical world. The robot no longer needs to be programmed to identify a soda can. It can use what it already knows from the internet to understand what a drink is, what it is for and how to grasp it, even if it has never seen that particular brand before.

However, training these models in the real world is very slow and, above all, dangerous. That is why researchers turned to an unexpected training ground: video games. Recent projects such as DeepMind's SIMA 2 (DeepMind) use complex virtual environments to train these intelligences before loading them into a robot. What is revolutionary about SIMA 2 is that it uses a massive model (Gemini) to reason before acting. If you ask it to "find a house the color of a tomato," the system does not search for random pixels. It reasons semantically that tomatoes are red and therefore it should look for a red structure. That ability to connect abstract instructions with visual perception and motor action is the missing link of robotics.

The logic of using simulations is hard to beat, and it revives Brooks's old principle: a physical robot learning to walk breaks and costs thousands of dollars to repair, while an agent in a video game can fall, crash and "die" a thousand times per second at no cost, learning causality, physics and consequences along the way.

We are moving from a Cartesian AI that separates mind from body to one that is truly embodied. The final goal is not for these systems to play video games or merely answer questions, but to develop the "situated cognition" needed to operate in our messy world. Researchers are beginning to find the body they were looking for to give artificial intelligence.

References and papers of interest

Keep reading