Scientists Build an AI That Imagines Paths to Goals — Just Like Your Brain Does

Scientists have built a new kind of artificial intelligence that can imagine different paths to a goal before actually moving — much like how a person pictures a route through a building before walking it. The system is called the Goal-Conditioned Memory Learner, or GCML. Researchers designed it to copy the way brain cells in animals create mental maps of their surroundings. The GCML can plan routes, avoid obstacles, and even solve tricky puzzles — all by running through imaginary journeys inside its memory.

Your brain has special cells called grid cells that fire in neat patterns as you move around. Scientists think these cells act like a built-in GPS, helping you keep track of where you are. The GCML copies this idea by using a math model of grid cells to build a 'cognitive map' — a kind of inner picture of a space. Once the map is built, the AI can think about how to get from one place to another without ever taking a single real step.

The AI also uses another type of brain cell called place cells. Place cells fire when an animal is in a specific spot, like a landmark on a map. In the GCML, place cells help the system remember what different locations look like. Together, grid cells and place cells give the AI a rich, detailed map it can explore in its imagination.

To plan a path, the GCML picks a starting point and a goal, then imagines small steps from start to finish. At each step, it chooses the action most likely to move it closer to the goal. A little bit of randomness is added on purpose so the AI explores more than just one path. This means it can come up with several different routes and then pick the best one.

The system can also dodge obstacles it has never seen before. When a new wall or barrier appears, the AI uses repulsive forces — like invisible bumpers — to steer away from the blockage. Amazingly, it does not need to relearn its whole map just because something new is in the way. This makes it much faster at adapting than many older AI systems.

The researchers also tested the GCML on an abstract puzzle called silhouette decomposition. In this puzzle, a shape made of puzzle pieces must be taken apart piece by piece until nothing is left. The GCML treated each removal like a move on a map, and it imagined sequences of removals until it found one that worked. Even though this task looks nothing like walking through a maze, the same mental-map idea worked just as well.

Older AI planning systems often need huge amounts of energy and computing power. The GCML is different because its cognitive map gives it a sense of direction right from the start. It heads toward the goal quickly, visiting far fewer dead-ends along the way. In real-life situations where speed matters, that advantage can be very important.

One of the best features of the GCML is that it learns while it explores. It does not need a teacher telling it the right answer every time. Instead, it watches what happens when it takes each action and slowly builds up its map and its sense of what moves are useful. Scientists call this self-supervised learning, and it is very similar to how children learn by playing and experimenting.

Because the learning rules are simple, the GCML could one day run on tiny, low-power computer chips. Researchers hope this will lead to smart devices that can think and plan without needing to connect to a big server. Imagine a small robot or a hearing aid that can figure out what to do next entirely on its own.

The scientists say the GCML is a big step toward AI that thinks more like a living brain. Future versions might be used in self-driving cars, medical robots, or space exploration drones. The team also plans to study whether real animals and people use the same kind of imagination-based planning that the GCML does. If they do, it would help scientists understand how our own minds work — and how to build even smarter machines inspired by nature.