Born wired: innate cortical connectivity plus local plasticity is enough to stand, walk and look
This paper demonstrates that a simulated quadruped can stand, walk toward visual targets, and exhibit gaze tracking using only innate cortical wiring and local correlation-based plasticity, without any global reward signals, gradients, or training loops.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Most of what we think of as "learning" in artificial intelligence involves a system trying millions of variations, failing, and slowly adjusting its internal settings until it finally gets the answer right. It is a process of trial and error, guided by a global goal, like a student memorizing a textbook by repeating chapters until the answers stick. But nature does not seem to work this way. A newborn animal does not start with a blank slate; it arrives with a body that knows how to stand and eyes that know how to track. This raises a fundamental question for scientists studying how brains work: how much of our behavior is truly learned, and how much is simply built into the wiring from the start? The prevailing view in modern machine learning suggests that complex behavior requires a massive amount of training data and a constant feedback loop to correct mistakes. However, a new line of inquiry asks if a brain could be born with a pre-drawn map of connections, where the only rule for change is a simple, local observation of which neurons fire together, without any need for a teacher or a global scorecard.
A researcher named Zhiwen Li has built a digital brain to test exactly this idea. Instead of training a system from scratch, they constructed a simulated brain with nearly 144,000 neurons and over 300,000 specific connections that were present from the very first moment of its existence. These initial connections, which the author calls an "instinct table," were not random; they were carefully designed to link sensory inputs, like seeing a red shape or feeling the body lying flat, directly to motor outputs, like the muscles needed to stand up or walk. Crucially, this system had no reward signal, no error correction, and no training loop. It was simply placed in a simulated world with a four-legged robot body and told to survive. The result was immediate and surprising: when placed on its belly, the robot stood up on its own. When shown a red object, it walked toward it. These actions were not learned through repetition; they were the direct result of the pre-wired connections firing in the correct sequence.
The experiment went further to prove that these behaviors were not accidental. When the researchers removed the sensory input entirely, the robot collapsed, proving that the behavior relied on the brain processing real-time information. When they wiped out the specific "instinct" connections but left the random background wiring intact, the robot could neither stand nor walk, confirming that the behavior came from the pre-specified map, not from chance. The study also showed that specific parts of the brain were responsible for specific actions. By silencing a large group of neurons in the prefrontal area—the part of the brain often associated with planning and decision-making—the researchers found that the robot could still react to a blue screen, but it completely lost the ability to walk toward a red object. This demonstrated that the "thought" of walking toward the red object was a specific pattern of activity in that group of cells, and without it, the behavior vanished, even though the basic reflexes remained.
Perhaps the most striking finding was that this system could learn a new skill during its lifetime without any external teacher. The researchers added a small group of cells that acted like a reward signal. When the robot was standing and saw a red object while hearing a specific sound, this signal fired. In that moment, the brain's local rule for change kicked in: it strengthened the connection between the cells that were active just before the sound and the cells that were active when the sound happened. After this single event, the robot learned to walk toward the red object just by hearing the sound, even when the object was not visible. This new ability was written into the brain's wiring in real-time, driven only by the coincidence of events and the presence of the reward signal, with no global objective function guiding the process.
However, the study also revealed the limits of this approach. While the robot could learn to walk, it struggled to stay upright for long periods when the reward signal was active for too long. The internal loops that allowed the brain to hold onto a thought or a plan also caused it to accumulate too much activity, eventually causing the robot to topple over. This suggests that while a pre-wired brain with simple local rules can generate complex behaviors and even learn new associations, it still needs a way to forget or reset, a mechanism that the current model lacks. The researchers also noted that the "instinct" table itself was built using data from a previously trained machine learning system, meaning the initial repertoire of movements was distilled from a learned policy before the brain was even created. This admission highlights that the system did not invent walking from nothing; it inherited a library of movements and learned how to trigger them in new contexts.
The work challenges the idea that intelligence requires a massive, top-down optimization process. Instead, it suggests that a significant portion of what an organism does is determined by the specific architecture it is born with. The brain is not a blank computer waiting for software; it is a machine with a hard-wired operating system that can be updated locally as it interacts with the world. The study shows that with the right starting connections, a system can stand, walk, and look without ever being told what a "correct" answer looks like. It does not claim to have solved the mystery of general intelligence, but it offers a concrete demonstration that complex, goal-directed behavior can emerge from a fixed set of rules and a simple, local learning mechanism, provided the initial wiring is correct. The findings suggest that the scarcity in biological intelligence may not be in the ability to learn, but in the initial structure that makes learning possible in the first place.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.